agora inbox for [email protected]
help / color / mirror / Atom feed[PATCH v9] Avoid orphaned objects dependencies
286+ messages / 2 participants
[nested] [flat]
* [PATCH v9] Avoid orphaned objects dependencies
@ 2024-03-29 15:43 Bertrand Drouvot <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Bertrand Drouvot @ 2024-03-29 15:43 UTC (permalink / raw)
It's currently possible to create orphaned objects dependencies, for example:
Scenario 1:
session 1: begin; drop schema schem;
session 2: create a function in the schema schem
session 1: commit;
With the above, the function created in session 2 would be linked to a non
existing schema.
Scenario 2:
session 1: begin; create a function in the schema schem
session 2: drop schema schem;
session 1: commit;
With the above, the function created in session 1 would be linked to a non
existing schema.
To avoid those scenarios, a new lock (that conflicts with a lock taken by DROP)
has been put in place when the dependencies are being recorded. With this in
place, the drop schema in scenario 2 would be locked.
Also, after the new lock attempt, the patch checks that the object still exists:
with this in place session 2 in scenario 1 would be locked and would report an
error once session 1 committs (that would not be the case should session 1 abort
the transaction).
If the object is dropped before the new lock attempt is triggered then the patch
would also report an error (but with less details).
The patch adds a few tests for some dependency cases (that would currently produce
orphaned objects):
- schema and function (as the above scenarios)
- alter a dependency (function and schema)
- function and arg type
- function and return type
- function and function
- domain and domain
- table and type
- server and foreign data wrapper
---
src/backend/catalog/aclchk.c | 1 +
src/backend/catalog/dependency.c | 83 +++++++++--
src/backend/catalog/heap.c | 7 +-
src/backend/catalog/index.c | 16 ++-
src/backend/catalog/objectaddress.c | 57 ++++++++
src/backend/catalog/pg_aggregate.c | 2 +-
src/backend/catalog/pg_attrdef.c | 1 +
src/backend/catalog/pg_cast.c | 2 +-
src/backend/catalog/pg_collation.c | 1 +
src/backend/catalog/pg_constraint.c | 10 +-
src/backend/catalog/pg_conversion.c | 2 +
src/backend/catalog/pg_depend.c | 34 ++++-
src/backend/catalog/pg_operator.c | 2 +-
src/backend/catalog/pg_proc.c | 12 +-
src/backend/catalog/pg_publication.c | 5 +
src/backend/catalog/pg_range.c | 3 +-
src/backend/catalog/pg_type.c | 18 ++-
src/backend/catalog/toasting.c | 1 +
src/backend/commands/alter.c | 3 +
src/backend/commands/amcmds.c | 1 +
src/backend/commands/cluster.c | 2 +
src/backend/commands/event_trigger.c | 1 +
src/backend/commands/extension.c | 4 +-
src/backend/commands/foreigncmds.c | 7 +
src/backend/commands/functioncmds.c | 3 +-
src/backend/commands/indexcmds.c | 2 +
src/backend/commands/opclasscmds.c | 18 +++
src/backend/commands/policy.c | 2 +
src/backend/commands/proclang.c | 2 +-
src/backend/commands/sequence.c | 1 +
src/backend/commands/statscmds.c | 3 +
src/backend/commands/tablecmds.c | 10 ++
src/backend/commands/trigger.c | 18 ++-
src/backend/commands/tsearchcmds.c | 8 +-
src/backend/commands/typecmds.c | 4 +
src/backend/rewrite/rewriteDefine.c | 1 +
src/backend/utils/errcodes.txt | 1 +
src/include/catalog/dependency.h | 9 +-
src/include/catalog/objectaddress.h | 1 +
.../expected/test_dependencies_locks.out | 129 ++++++++++++++++++
src/test/isolation/isolation_schedule | 1 +
.../specs/test_dependencies_locks.spec | 89 ++++++++++++
42 files changed, 534 insertions(+), 43 deletions(-)
34.8% src/backend/catalog/
16.2% src/backend/commands/
28.0% src/test/isolation/expected/
17.4% src/test/isolation/specs/
3.3% src/
diff --git a/src/backend/catalog/aclchk.c b/src/backend/catalog/aclchk.c
index 143876b77f..6b91402a9f 100644
--- a/src/backend/catalog/aclchk.c
+++ b/src/backend/catalog/aclchk.c
@@ -1413,6 +1413,7 @@ SetDefaultACL(InternalDefaultACL *iacls)
referenced.objectId = iacls->nspid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
}
}
diff --git a/src/backend/catalog/dependency.c b/src/backend/catalog/dependency.c
index d4b5b2ade1..42293220fd 100644
--- a/src/backend/catalog/dependency.c
+++ b/src/backend/catalog/dependency.c
@@ -1517,6 +1517,68 @@ AcquireDeletionLock(const ObjectAddress *object, int flags)
}
}
+/*
+ * depLockAndCheckObject
+ *
+ * Lock the object that we are about to record a dependency on.
+ * After it's locked, verify that it hasn't been dropped while we
+ * weren't looking. If the object has been dropped, this function
+ * does not return!
+ */
+void
+depLockAndCheckObject(const ObjectAddress *object)
+{
+ char *object_description;
+
+ if (isObjectPinned(object))
+ return;
+
+ /*
+ * Those don't rely on LockDatabaseObject() when being dropped (see
+ * AcquireDeletionLock()). Also it looks like they can not produce
+ * orphaned dependent objects when being dropped.
+ */
+ if (object->classId == RelationRelationId || object->classId == AuthMemRelationId)
+ return;
+
+ object_description = getObjectDescription(object, true);
+
+ /* assume we should lock the whole object not a sub-object */
+ LockDatabaseObject(object->classId, object->objectId, 0, AccessShareLock);
+
+ /* check if object still exists */
+ if (!ObjectByIdExist(object))
+ {
+ if (object_description)
+ ereport(ERROR,
+ (errcode(ERRCODE_DEPENDENT_OBJECTS_DOES_NOT_EXIST),
+ errmsg("%s does not exist", object_description)));
+ else
+ ereport(ERROR,
+ (errcode(ERRCODE_DEPENDENT_OBJECTS_DOES_NOT_EXIST),
+ errmsg("a dependent object does not exist")));
+ }
+
+ if (object_description)
+ pfree(object_description);
+
+ return;
+}
+
+void
+depLockAndCheckObjects(const ObjectAddress *object, int nobject)
+{
+ int i;
+
+ if (nobject < 0)
+ return;
+
+ for (i = 0; i < nobject; i++, object++)
+ depLockAndCheckObject(object);
+
+ return;
+}
+
/*
* ReleaseDeletionLock - release an object deletion lock
*
@@ -1562,13 +1624,8 @@ recordDependencyOnExpr(const ObjectAddress *depender,
/* Scan the expression tree for referenceable objects */
find_expr_references_walker(expr, &context);
- /* Remove any duplicates */
- eliminate_duplicate_dependencies(context.addrs);
-
- /* And record 'em */
- recordMultipleDependencies(depender,
- context.addrs->refs, context.addrs->numrefs,
- behavior);
+ /* Record all of them (this includes duplicate elimination) */
+ lock_record_object_address_dependencies(depender, context.addrs, behavior);
free_object_addresses(context.addrs);
}
@@ -1652,9 +1709,12 @@ recordDependencyOnSingleRelExpr(const ObjectAddress *depender,
/* Record the self-dependencies with the appropriate direction */
if (!reverse_self)
+ {
+ depLockAndCheckObjects(self_addrs->refs, self_addrs->numrefs);
recordMultipleDependencies(depender,
self_addrs->refs, self_addrs->numrefs,
self_behavior);
+ }
else
{
/* Can't use recordMultipleDependencies, so do it the hard way */
@@ -1664,6 +1724,7 @@ recordDependencyOnSingleRelExpr(const ObjectAddress *depender,
{
ObjectAddress *thisobj = self_addrs->refs + selfref;
+ depLockAndCheckObject(depender);
recordDependencyOn(thisobj, depender, self_behavior);
}
}
@@ -1672,6 +1733,7 @@ recordDependencyOnSingleRelExpr(const ObjectAddress *depender,
}
/* Record the external dependencies */
+ depLockAndCheckObjects(context.addrs->refs, context.addrs->numrefs);
recordMultipleDependencies(depender,
context.addrs->refs, context.addrs->numrefs,
behavior);
@@ -2737,11 +2799,12 @@ stack_address_present_add_flags(const ObjectAddress *object,
* removing any duplicates.
*/
void
-record_object_address_dependencies(const ObjectAddress *depender,
- ObjectAddresses *referenced,
- DependencyType behavior)
+lock_record_object_address_dependencies(const ObjectAddress *depender,
+ ObjectAddresses *referenced,
+ DependencyType behavior)
{
eliminate_duplicate_dependencies(referenced);
+ depLockAndCheckObjects(referenced->refs, referenced->numrefs);
recordMultipleDependencies(depender,
referenced->refs, referenced->numrefs,
behavior);
diff --git a/src/backend/catalog/heap.c b/src/backend/catalog/heap.c
index a122bbffce..113b275fe1 100644
--- a/src/backend/catalog/heap.c
+++ b/src/backend/catalog/heap.c
@@ -843,6 +843,7 @@ AddNewAttributeTuples(Oid new_rel_oid,
ObjectAddressSubSet(myself, RelationRelationId, new_rel_oid, i + 1);
ObjectAddressSet(referenced, TypeRelationId,
tupdesc->attrs[i].atttypid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* The default collation is pinned, so don't bother recording it */
@@ -851,6 +852,7 @@ AddNewAttributeTuples(Oid new_rel_oid,
{
ObjectAddressSet(referenced, CollationRelationId,
tupdesc->attrs[i].attcollation);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
}
@@ -1471,7 +1473,7 @@ heap_create_with_catalog(const char *relname,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
}
@@ -3393,7 +3395,7 @@ StorePartitionKey(Relation rel,
}
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
/*
@@ -3409,6 +3411,7 @@ StorePartitionKey(Relation rel,
ObjectAddressSubSet(referenced, RelationRelationId,
RelationGetRelid(rel), partattrs[i]);
+ depLockAndCheckObject(&myself);
recordDependencyOn(&referenced, &myself, DEPENDENCY_INTERNAL);
}
diff --git a/src/backend/catalog/index.c b/src/backend/catalog/index.c
index 55fdde4b24..26201e7610 100644
--- a/src/backend/catalog/index.c
+++ b/src/backend/catalog/index.c
@@ -1144,7 +1144,7 @@ index_create(Relation heapRelation,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_AUTO);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_AUTO);
free_object_addresses(addrs);
}
@@ -1157,9 +1157,11 @@ index_create(Relation heapRelation,
if (OidIsValid(parentIndexRelid))
{
ObjectAddressSet(referenced, RelationRelationId, parentIndexRelid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_PARTITION_PRI);
ObjectAddressSet(referenced, RelationRelationId, heapRelationId);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_PARTITION_SEC);
}
@@ -1185,7 +1187,7 @@ index_create(Relation heapRelation,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
/* Store dependencies on anything mentioned in index expressions */
@@ -1987,6 +1989,14 @@ index_constraint_create(Relation heapRelation,
*/
ObjectAddressSet(myself, ConstraintRelationId, conOid);
ObjectAddressSet(idxaddr, RelationRelationId, indexRelationId);
+
+ /*
+ * CommandCounterIncrement here to ensure the new constraint entry is
+ * visible when we'll check of object existence when recording the
+ * dependencies.
+ */
+ CommandCounterIncrement();
+ depLockAndCheckObject(&myself);
recordDependencyOn(&idxaddr, &myself, DEPENDENCY_INTERNAL);
/*
@@ -1998,9 +2008,11 @@ index_constraint_create(Relation heapRelation,
ObjectAddress referenced;
ObjectAddressSet(referenced, ConstraintRelationId, parentConstraintId);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_PARTITION_PRI);
ObjectAddressSet(referenced, RelationRelationId,
RelationGetRelid(heapRelation));
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_PARTITION_SEC);
}
diff --git a/src/backend/catalog/objectaddress.c b/src/backend/catalog/objectaddress.c
index 7b536ac6fd..6d7abd3738 100644
--- a/src/backend/catalog/objectaddress.c
+++ b/src/backend/catalog/objectaddress.c
@@ -2590,6 +2590,63 @@ get_object_namespace(const ObjectAddress *address)
return oid;
}
+/*
+ * ObjectByIdExist
+ *
+ * Return whether the given object exists.
+ *
+ * Works for most catalogs, if no special processing is needed.
+ */
+bool
+ObjectByIdExist(const ObjectAddress *address)
+{
+ HeapTuple tuple;
+ int cache;
+ const ObjectPropertyType *property;
+
+ property = get_object_property_data(address->classId);
+
+ cache = property->oid_catcache_id;
+
+ if (cache >= 0)
+ {
+ /* Fetch tuple from syscache. */
+ tuple = SearchSysCache1(cache, ObjectIdGetDatum(address->objectId));
+
+ if (!HeapTupleIsValid(tuple))
+ {
+ return false;
+ }
+
+ ReleaseSysCache(tuple);
+
+ return true;
+ }
+ else
+ {
+ Relation rel;
+ ScanKeyData skey[1];
+ SysScanDesc scan;
+
+ rel = table_open(address->classId, AccessShareLock);
+
+ ScanKeyInit(&skey[0],
+ get_object_attnum_oid(address->classId),
+ BTEqualStrategyNumber, F_OIDEQ,
+ ObjectIdGetDatum(address->objectId));
+
+ scan = systable_beginscan(rel, get_object_oid_index(address->classId), true,
+ NULL, 1, skey);
+
+ /* we expect exactly one match */
+ tuple = systable_getnext(scan);
+ systable_endscan(scan);
+ table_close(rel, AccessShareLock);
+
+ return (HeapTupleIsValid(tuple));
+ }
+}
+
/*
* Return ObjectType for the given object type as given by
* getObjectTypeDescription; if no valid ObjectType code exists, but it's a
diff --git a/src/backend/catalog/pg_aggregate.c b/src/backend/catalog/pg_aggregate.c
index 90fc7db949..9b5c8c8fb0 100644
--- a/src/backend/catalog/pg_aggregate.c
+++ b/src/backend/catalog/pg_aggregate.c
@@ -805,7 +805,7 @@ AggregateCreate(const char *aggName,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
return myself;
}
diff --git a/src/backend/catalog/pg_attrdef.c b/src/backend/catalog/pg_attrdef.c
index 003ae70b4d..47b7fe75ad 100644
--- a/src/backend/catalog/pg_attrdef.c
+++ b/src/backend/catalog/pg_attrdef.c
@@ -178,6 +178,7 @@ StoreAttrDefault(Relation rel, AttrNumber attnum,
colobject.objectId = RelationGetRelid(rel);
colobject.objectSubId = attnum;
+ depLockAndCheckObject(&colobject);
recordDependencyOn(&defobject, &colobject,
attgenerated ? DEPENDENCY_INTERNAL : DEPENDENCY_AUTO);
diff --git a/src/backend/catalog/pg_cast.c b/src/backend/catalog/pg_cast.c
index 5a5b855d51..2593bb3634 100644
--- a/src/backend/catalog/pg_cast.c
+++ b/src/backend/catalog/pg_cast.c
@@ -121,7 +121,7 @@ CastCreate(Oid sourcetypeid, Oid targettypeid,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, behavior);
+ lock_record_object_address_dependencies(&myself, addrs, behavior);
free_object_addresses(addrs);
/* dependency on extension */
diff --git a/src/backend/catalog/pg_collation.c b/src/backend/catalog/pg_collation.c
index 7f2f701229..43ceb426d4 100644
--- a/src/backend/catalog/pg_collation.c
+++ b/src/backend/catalog/pg_collation.c
@@ -218,6 +218,7 @@ CollationCreate(const char *collname, Oid collnamespace,
referenced.classId = NamespaceRelationId;
referenced.objectId = collnamespace;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* create dependency on owner */
diff --git a/src/backend/catalog/pg_constraint.c b/src/backend/catalog/pg_constraint.c
index 3baf9231ed..91af823ee5 100644
--- a/src/backend/catalog/pg_constraint.c
+++ b/src/backend/catalog/pg_constraint.c
@@ -277,8 +277,8 @@ CreateConstraintEntry(const char *constraintName,
add_exact_object_address(&domobject, addrs_auto);
}
- record_object_address_dependencies(&conobject, addrs_auto,
- DEPENDENCY_AUTO);
+ lock_record_object_address_dependencies(&conobject, addrs_auto,
+ DEPENDENCY_AUTO);
free_object_addresses(addrs_auto);
/* Handle set of normal dependencies */
@@ -352,8 +352,8 @@ CreateConstraintEntry(const char *constraintName,
}
}
- record_object_address_dependencies(&conobject, addrs_normal,
- DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&conobject, addrs_normal,
+ DEPENDENCY_NORMAL);
free_object_addresses(addrs_normal);
/*
@@ -858,9 +858,11 @@ ConstraintSetParentConstraint(Oid childConstrId,
ObjectAddressSet(depender, ConstraintRelationId, childConstrId);
ObjectAddressSet(referenced, ConstraintRelationId, parentConstrId);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&depender, &referenced, DEPENDENCY_PARTITION_PRI);
ObjectAddressSet(referenced, RelationRelationId, childTableId);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&depender, &referenced, DEPENDENCY_PARTITION_SEC);
}
else
diff --git a/src/backend/catalog/pg_conversion.c b/src/backend/catalog/pg_conversion.c
index 0770878eac..479dd528eb 100644
--- a/src/backend/catalog/pg_conversion.c
+++ b/src/backend/catalog/pg_conversion.c
@@ -116,12 +116,14 @@ ConversionCreate(const char *conname, Oid connamespace,
referenced.classId = ProcedureRelationId;
referenced.objectId = conproc;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* create dependency on namespace */
referenced.classId = NamespaceRelationId;
referenced.objectId = connamespace;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* create dependency on owner */
diff --git a/src/backend/catalog/pg_depend.c b/src/backend/catalog/pg_depend.c
index cfd7ef51df..47968ee8c8 100644
--- a/src/backend/catalog/pg_depend.c
+++ b/src/backend/catalog/pg_depend.c
@@ -20,21 +20,20 @@
#include "catalog/catalog.h"
#include "catalog/dependency.h"
#include "catalog/indexing.h"
+#include "catalog/pg_auth_members.h"
#include "catalog/pg_constraint.h"
#include "catalog/pg_depend.h"
#include "catalog/pg_extension.h"
#include "catalog/partition.h"
#include "commands/extension.h"
#include "miscadmin.h"
+#include "storage/lock.h"
#include "utils/fmgroids.h"
#include "utils/lsyscache.h"
#include "utils/syscache.h"
#include "utils/rel.h"
-static bool isObjectPinned(const ObjectAddress *object);
-
-
/*
* Record a dependency between 2 objects via their respective objectAddress.
* The first argument is the dependent object, the second the one it
@@ -100,6 +99,25 @@ recordMultipleDependencies(const ObjectAddress *depender,
slot_init_count = 0;
for (i = 0; i < nreferenced; i++, referenced++)
{
+#ifdef USE_ASSERT_CHECKING
+ LOCKTAG tag;
+
+ SET_LOCKTAG_OBJECT(tag,
+ MyDatabaseId,
+ referenced->classId,
+ referenced->objectId,
+ 0);
+
+ /*
+ * Assert the referenced object is locked (see
+ * depLockAndCheckObject()) when recording the dependency.
+ */
+ Assert(isObjectPinned(referenced) ||
+ referenced->classId == RelationRelationId ||
+ referenced->classId == AuthMemRelationId ||
+ LockHeldByMe(&tag, AccessShareLock));
+#endif
+
/*
* If the referenced object is pinned by the system, there's no real
* need to record dependencies on it. This saves lots of space in
@@ -239,6 +257,7 @@ recordDependencyOnCurrentExtension(const ObjectAddress *object,
extension.objectId = CurrentExtensionObject;
extension.objectSubId = 0;
+ depLockAndCheckObject(&extension);
recordDependencyOn(object, &extension, DEPENDENCY_EXTENSION);
}
}
@@ -501,11 +520,18 @@ changeDependencyFor(Oid classId, Oid objectId,
depAddr.classId = classId;
depAddr.objectId = objectId;
depAddr.objectSubId = 0;
+ depLockAndCheckObject(&objAddr);
recordDependencyOn(&depAddr, &objAddr, DEPENDENCY_NORMAL);
return 1;
}
+ /*
+ * Acquire a lock and check object still exists while changing the
+ * dependency.
+ */
+ depLockAndCheckObject(&objAddr);
+
depRel = table_open(DependRelationId, RowExclusiveLock);
/* There should be existing dependency record(s), so search. */
@@ -706,7 +732,7 @@ changeDependenciesOn(Oid refClassId, Oid oldRefObjectId,
* The passed subId, if any, is ignored; we assume that only whole objects
* are pinned (and that this implies pinning their components).
*/
-static bool
+bool
isObjectPinned(const ObjectAddress *object)
{
return IsPinnedObject(object->classId, object->objectId);
diff --git a/src/backend/catalog/pg_operator.c b/src/backend/catalog/pg_operator.c
index 65b45a424a..7e40471baf 100644
--- a/src/backend/catalog/pg_operator.c
+++ b/src/backend/catalog/pg_operator.c
@@ -930,7 +930,7 @@ makeOperatorDependencies(HeapTuple tuple,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
/* Dependency on owner */
diff --git a/src/backend/catalog/pg_proc.c b/src/backend/catalog/pg_proc.c
index 528c17cd7f..aa36a07648 100644
--- a/src/backend/catalog/pg_proc.c
+++ b/src/backend/catalog/pg_proc.c
@@ -593,6 +593,13 @@ ProcedureCreate(const char *procedureName,
if (is_update)
deleteDependencyRecordsFor(ProcedureRelationId, retval, true);
+ /*
+ * CommandCounterIncrement here to ensure the new function entry is
+ * visible when we'll check of object existence when recording the
+ * dependencies.
+ */
+ CommandCounterIncrement();
+
addrs = new_object_addresses();
ObjectAddressSet(myself, ProcedureRelationId, retval);
@@ -637,7 +644,7 @@ ProcedureCreate(const char *procedureName,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
/* dependency on SQL routine body */
@@ -674,9 +681,6 @@ ProcedureCreate(const char *procedureName,
ArrayType *set_items = NULL;
int save_nestlevel = 0;
- /* Advance command counter so new tuple can be seen by validator */
- CommandCounterIncrement();
-
/*
* Set per-function configuration parameters so that the validation is
* done with the environment the function expects. However, if
diff --git a/src/backend/catalog/pg_publication.c b/src/backend/catalog/pg_publication.c
index 0602398a54..6654b65f41 100644
--- a/src/backend/catalog/pg_publication.c
+++ b/src/backend/catalog/pg_publication.c
@@ -438,10 +438,12 @@ publication_add_relation(Oid pubid, PublicationRelInfo *pri,
/* Add dependency on the publication */
ObjectAddressSet(referenced, PublicationRelationId, pubid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
/* Add dependency on the relation */
ObjectAddressSet(referenced, RelationRelationId, relid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
/* Add dependency on the objects mentioned in the qualifications */
@@ -454,6 +456,7 @@ publication_add_relation(Oid pubid, PublicationRelInfo *pri,
for (int i = 0; i < natts; i++)
{
ObjectAddressSubSet(referenced, RelationRelationId, relid, attarray[i]);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
@@ -661,10 +664,12 @@ publication_add_schema(Oid pubid, Oid schemaid, bool if_not_exists)
/* Add dependency on the publication */
ObjectAddressSet(referenced, PublicationRelationId, pubid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
/* Add dependency on the schema */
ObjectAddressSet(referenced, NamespaceRelationId, schemaid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
/* Close the table */
diff --git a/src/backend/catalog/pg_range.c b/src/backend/catalog/pg_range.c
index 501a6ba410..a2f343a81a 100644
--- a/src/backend/catalog/pg_range.c
+++ b/src/backend/catalog/pg_range.c
@@ -92,13 +92,14 @@ RangeCreate(Oid rangeTypeOid, Oid rangeSubType, Oid rangeCollation,
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
/* record multirange type's dependency on the range type */
referencing.classId = TypeRelationId;
referencing.objectId = multirangeTypeOid;
referencing.objectSubId = 0;
+ depLockAndCheckObject(&myself);
recordDependencyOn(&referencing, &myself, DEPENDENCY_INTERNAL);
table_close(pg_range, RowExclusiveLock);
diff --git a/src/backend/catalog/pg_type.c b/src/backend/catalog/pg_type.c
index 395dec8ed8..91139e275f 100644
--- a/src/backend/catalog/pg_type.c
+++ b/src/backend/catalog/pg_type.c
@@ -494,6 +494,14 @@ TypeCreate(Oid newTypeOid,
* Create dependencies. We can/must skip this in bootstrap mode.
*/
if (!IsBootstrapProcessingMode())
+ {
+ /*
+ * CommandCounterIncrement here to ensure the new type entry is
+ * visible when we'll check of object existence when recording the
+ * dependencies.
+ */
+ CommandCounterIncrement();
+
GenerateTypeDependencies(tup,
pg_type_desc,
(defaultTypeBin ?
@@ -505,6 +513,7 @@ TypeCreate(Oid newTypeOid,
isDependentType,
true, /* make extension dependency */
rebuildDeps);
+ }
/* Post creation hook for new type */
InvokeObjectPostCreateHook(TypeRelationId, typeObjectId, 0);
@@ -703,7 +712,7 @@ GenerateTypeDependencies(HeapTuple typeTuple,
add_exact_object_address(&referenced, addrs_normal);
}
- record_object_address_dependencies(&myself, addrs_normal, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs_normal, DEPENDENCY_NORMAL);
free_object_addresses(addrs_normal);
/* Normal dependency on the default expression. */
@@ -724,9 +733,15 @@ GenerateTypeDependencies(HeapTuple typeTuple,
ObjectAddressSet(referenced, RelationRelationId, typeForm->typrelid);
if (relationKind != RELKIND_COMPOSITE_TYPE)
+ {
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_INTERNAL);
+ }
else
+ {
+ depLockAndCheckObject(&myself);
recordDependencyOn(&referenced, &myself, DEPENDENCY_INTERNAL);
+ }
}
/*
@@ -737,6 +752,7 @@ GenerateTypeDependencies(HeapTuple typeTuple,
if (OidIsValid(typeForm->typelem))
{
ObjectAddressSet(referenced, TypeRelationId, typeForm->typelem);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced,
isImplicitArray ? DEPENDENCY_INTERNAL : DEPENDENCY_NORMAL);
}
diff --git a/src/backend/catalog/toasting.c b/src/backend/catalog/toasting.c
index 738bc46ae8..a471833efe 100644
--- a/src/backend/catalog/toasting.c
+++ b/src/backend/catalog/toasting.c
@@ -367,6 +367,7 @@ create_toast_table(Relation rel, Oid toastOid, Oid toastIndexOid,
toastobject.objectId = toast_relid;
toastobject.objectSubId = 0;
+ depLockAndCheckObject(&baseobject);
recordDependencyOn(&toastobject, &baseobject, DEPENDENCY_INTERNAL);
}
diff --git a/src/backend/commands/alter.c b/src/backend/commands/alter.c
index 4f99ebb447..e3ad5db529 100644
--- a/src/backend/commands/alter.c
+++ b/src/backend/commands/alter.c
@@ -501,7 +501,10 @@ ExecAlterObjectDependsStmt(AlterObjectDependsStmt *stmt, ObjectAddress *refAddre
currexts = getAutoExtensionsOfObject(address.classId,
address.objectId);
if (!list_member_oid(currexts, refAddr.objectId))
+ {
+ depLockAndCheckObject(&refAddr);
recordDependencyOn(&address, &refAddr, DEPENDENCY_AUTO_EXTENSION);
+ }
}
return address;
diff --git a/src/backend/commands/amcmds.c b/src/backend/commands/amcmds.c
index aaa0f9a1dc..0b4e44131a 100644
--- a/src/backend/commands/amcmds.c
+++ b/src/backend/commands/amcmds.c
@@ -104,6 +104,7 @@ CreateAccessMethod(CreateAmStmt *stmt)
referenced.objectId = amhandler;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
recordDependencyOnCurrentExtension(&myself, false);
diff --git a/src/backend/commands/cluster.c b/src/backend/commands/cluster.c
index 78f96789b0..73bcd68a1f 100644
--- a/src/backend/commands/cluster.c
+++ b/src/backend/commands/cluster.c
@@ -1381,6 +1381,7 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class,
{
baseobject.objectId = r1;
toastobject.objectId = relform1->reltoastrelid;
+ depLockAndCheckObject(&baseobject);
recordDependencyOn(&toastobject, &baseobject,
DEPENDENCY_INTERNAL);
}
@@ -1389,6 +1390,7 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class,
{
baseobject.objectId = r2;
toastobject.objectId = relform2->reltoastrelid;
+ depLockAndCheckObject(&baseobject);
recordDependencyOn(&toastobject, &baseobject,
DEPENDENCY_INTERNAL);
}
diff --git a/src/backend/commands/event_trigger.c b/src/backend/commands/event_trigger.c
index 7a5ed6b985..ddcba91c92 100644
--- a/src/backend/commands/event_trigger.c
+++ b/src/backend/commands/event_trigger.c
@@ -327,6 +327,7 @@ insert_event_trigger_tuple(const char *trigname, const char *eventname, Oid evtO
referenced.classId = ProcedureRelationId;
referenced.objectId = funcoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* Depend on extension, if any. */
diff --git a/src/backend/commands/extension.c b/src/backend/commands/extension.c
index 1643c8c69a..fdcffcea01 100644
--- a/src/backend/commands/extension.c
+++ b/src/backend/commands/extension.c
@@ -1935,7 +1935,7 @@ InsertExtensionTuple(const char *extName, Oid extOwner,
}
/* Record all of them (this includes duplicate elimination) */
- record_object_address_dependencies(&myself, refobjs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, refobjs, DEPENDENCY_NORMAL);
free_object_addresses(refobjs);
/* Post creation hook for new extension */
@@ -3258,6 +3258,7 @@ ApplyExtensionUpdates(Oid extensionOid,
otherext.objectId = reqext;
otherext.objectSubId = 0;
+ depLockAndCheckObject(&otherext);
recordDependencyOn(&myself, &otherext, DEPENDENCY_NORMAL);
}
@@ -3414,6 +3415,7 @@ ExecAlterExtensionContentsRecurse(AlterExtensionContentsStmt *stmt,
/*
* OK, add the dependency.
*/
+ depLockAndCheckObject(&extension);
recordDependencyOn(&object, &extension, DEPENDENCY_EXTENSION);
/*
diff --git a/src/backend/commands/foreigncmds.c b/src/backend/commands/foreigncmds.c
index cf61bbac1f..5e98f3d0d6 100644
--- a/src/backend/commands/foreigncmds.c
+++ b/src/backend/commands/foreigncmds.c
@@ -642,6 +642,7 @@ CreateForeignDataWrapper(ParseState *pstate, CreateFdwStmt *stmt)
referenced.classId = ProcedureRelationId;
referenced.objectId = fdwhandler;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
@@ -650,6 +651,7 @@ CreateForeignDataWrapper(ParseState *pstate, CreateFdwStmt *stmt)
referenced.classId = ProcedureRelationId;
referenced.objectId = fdwvalidator;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
@@ -811,6 +813,7 @@ AlterForeignDataWrapper(ParseState *pstate, AlterFdwStmt *stmt)
referenced.classId = ProcedureRelationId;
referenced.objectId = fdwhandler;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
@@ -819,6 +822,7 @@ AlterForeignDataWrapper(ParseState *pstate, AlterFdwStmt *stmt)
referenced.classId = ProcedureRelationId;
referenced.objectId = fdwvalidator;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
}
@@ -951,6 +955,7 @@ CreateForeignServer(CreateForeignServerStmt *stmt)
referenced.classId = ForeignDataWrapperRelationId;
referenced.objectId = fdw->fdwid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
recordDependencyOnOwner(ForeignServerRelationId, srvId, ownerId);
@@ -1195,6 +1200,7 @@ CreateUserMapping(CreateUserMappingStmt *stmt)
referenced.classId = ForeignServerRelationId;
referenced.objectId = srv->serverid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
if (OidIsValid(useId))
@@ -1472,6 +1478,7 @@ CreateForeignTable(CreateForeignTableStmt *stmt, Oid relid)
referenced.classId = ForeignServerRelationId;
referenced.objectId = server->serverid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
table_close(ftrel, RowExclusiveLock);
diff --git a/src/backend/commands/functioncmds.c b/src/backend/commands/functioncmds.c
index 6593fd7d81..2a378a53d8 100644
--- a/src/backend/commands/functioncmds.c
+++ b/src/backend/commands/functioncmds.c
@@ -1459,6 +1459,7 @@ AlterFunction(ParseState *pstate, AlterFunctionStmt *stmt)
referenced.classId = ProcedureRelationId;
referenced.objectId = newsupport;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&address, &referenced, DEPENDENCY_NORMAL);
}
@@ -1979,7 +1980,7 @@ CreateTransform(CreateTransformStmt *stmt)
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
/* dependency on extension */
diff --git a/src/backend/commands/indexcmds.c b/src/backend/commands/indexcmds.c
index 309389e20d..487708d885 100644
--- a/src/backend/commands/indexcmds.c
+++ b/src/backend/commands/indexcmds.c
@@ -4377,8 +4377,10 @@ IndexSetParentIndex(Relation partitionIdx, Oid parentOid)
ObjectAddressSet(parentIdx, RelationRelationId, parentOid);
ObjectAddressSet(partitionTbl, RelationRelationId,
partitionIdx->rd_index->indrelid);
+ depLockAndCheckObject(&parentIdx);
recordDependencyOn(&partIdx, &parentIdx,
DEPENDENCY_PARTITION_PRI);
+ depLockAndCheckObject(&partitionTbl);
recordDependencyOn(&partIdx, &partitionTbl,
DEPENDENCY_PARTITION_SEC);
}
diff --git a/src/backend/commands/opclasscmds.c b/src/backend/commands/opclasscmds.c
index b8b5c147c5..e7b1cb911e 100644
--- a/src/backend/commands/opclasscmds.c
+++ b/src/backend/commands/opclasscmds.c
@@ -298,12 +298,14 @@ CreateOpFamily(CreateOpFamilyStmt *stmt, const char *opfname,
referenced.classId = AccessMethodRelationId;
referenced.objectId = amoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
/* dependency on namespace */
referenced.classId = NamespaceRelationId;
referenced.objectId = namespaceoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* dependency on owner */
@@ -725,18 +727,21 @@ DefineOpClass(CreateOpClassStmt *stmt)
referenced.classId = NamespaceRelationId;
referenced.objectId = namespaceoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* dependency on opfamily */
referenced.classId = OperatorFamilyRelationId;
referenced.objectId = opfamilyoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
/* dependency on indexed datatype */
referenced.classId = TypeRelationId;
referenced.objectId = typeoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
/* dependency on storage datatype */
@@ -745,6 +750,7 @@ DefineOpClass(CreateOpClassStmt *stmt)
referenced.classId = TypeRelationId;
referenced.objectId = storageoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
@@ -1486,6 +1492,13 @@ storeOperators(List *opfamilyname, Oid amoid, Oid opfamilyoid,
heap_freetuple(tup);
+ /*
+ * CommandCounterIncrement here to ensure the new operator entry is
+ * visible when we'll check of object existence when recording the
+ * dependencies.
+ */
+ CommandCounterIncrement();
+
/* Make its dependencies */
myself.classId = AccessMethodOperatorRelationId;
myself.objectId = entryoid;
@@ -1496,6 +1509,7 @@ storeOperators(List *opfamilyname, Oid amoid, Oid opfamilyoid,
referenced.objectSubId = 0;
/* see comments in amapi.h about dependency strength */
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced,
op->ref_is_hard ? DEPENDENCY_NORMAL : DEPENDENCY_AUTO);
@@ -1504,6 +1518,7 @@ storeOperators(List *opfamilyname, Oid amoid, Oid opfamilyoid,
referenced.objectId = op->refobjid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced,
op->ref_is_hard ? DEPENDENCY_INTERNAL : DEPENDENCY_AUTO);
@@ -1514,6 +1529,7 @@ storeOperators(List *opfamilyname, Oid amoid, Oid opfamilyoid,
referenced.objectId = op->sortfamily;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced,
op->ref_is_hard ? DEPENDENCY_NORMAL : DEPENDENCY_AUTO);
}
@@ -1597,6 +1613,7 @@ storeProcedures(List *opfamilyname, Oid amoid, Oid opfamilyoid,
referenced.objectSubId = 0;
/* see comments in amapi.h about dependency strength */
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced,
proc->ref_is_hard ? DEPENDENCY_NORMAL : DEPENDENCY_AUTO);
@@ -1605,6 +1622,7 @@ storeProcedures(List *opfamilyname, Oid amoid, Oid opfamilyoid,
referenced.objectId = proc->refobjid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced,
proc->ref_is_hard ? DEPENDENCY_INTERNAL : DEPENDENCY_AUTO);
diff --git a/src/backend/commands/policy.c b/src/backend/commands/policy.c
index 6ff3eba824..5a50ad73fb 100644
--- a/src/backend/commands/policy.c
+++ b/src/backend/commands/policy.c
@@ -722,6 +722,7 @@ CreatePolicy(CreatePolicyStmt *stmt)
myself.objectId = policy_id;
myself.objectSubId = 0;
+ depLockAndCheckObject(&target);
recordDependencyOn(&myself, &target, DEPENDENCY_AUTO);
recordDependencyOnExpr(&myself, qual, qual_pstate->p_rtable,
@@ -1053,6 +1054,7 @@ AlterPolicy(AlterPolicyStmt *stmt)
myself.objectId = policy_id;
myself.objectSubId = 0;
+ depLockAndCheckObject(&target);
recordDependencyOn(&myself, &target, DEPENDENCY_AUTO);
recordDependencyOnExpr(&myself, qual, qual_parse_rtable, DEPENDENCY_NORMAL);
diff --git a/src/backend/commands/proclang.c b/src/backend/commands/proclang.c
index 881f90017e..679155162e 100644
--- a/src/backend/commands/proclang.c
+++ b/src/backend/commands/proclang.c
@@ -205,7 +205,7 @@ CreateProceduralLanguage(CreatePLangStmt *stmt)
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
/* Post creation hook for new procedural language */
diff --git a/src/backend/commands/sequence.c b/src/backend/commands/sequence.c
index 28f8522264..0d37576d22 100644
--- a/src/backend/commands/sequence.c
+++ b/src/backend/commands/sequence.c
@@ -1681,6 +1681,7 @@ process_owned_by(Relation seqrel, List *owned_by, bool for_identity)
depobject.classId = RelationRelationId;
depobject.objectId = RelationGetRelid(seqrel);
depobject.objectSubId = 0;
+ depLockAndCheckObject(&refobject);
recordDependencyOn(&depobject, &refobject, deptype);
}
diff --git a/src/backend/commands/statscmds.c b/src/backend/commands/statscmds.c
index 1db3ef69d2..455ec046f8 100644
--- a/src/backend/commands/statscmds.c
+++ b/src/backend/commands/statscmds.c
@@ -536,6 +536,7 @@ CreateStatistics(CreateStatsStmt *stmt)
for (i = 0; i < nattnums; i++)
{
ObjectAddressSubSet(parentobject, RelationRelationId, relid, attnums[i]);
+ depLockAndCheckObject(&parentobject);
recordDependencyOn(&myself, &parentobject, DEPENDENCY_AUTO);
}
@@ -553,6 +554,7 @@ CreateStatistics(CreateStatsStmt *stmt)
if (!nattnums)
{
ObjectAddressSet(parentobject, RelationRelationId, relid);
+ depLockAndCheckObject(&parentobject);
recordDependencyOn(&myself, &parentobject, DEPENDENCY_AUTO);
}
@@ -573,6 +575,7 @@ CreateStatistics(CreateStatsStmt *stmt)
* than the underlying table(s).
*/
ObjectAddressSet(parentobject, NamespaceRelationId, namespaceId);
+ depLockAndCheckObject(&parentobject);
recordDependencyOn(&myself, &parentobject, DEPENDENCY_NORMAL);
recordDependencyOnOwner(StatisticExtRelationId, statoid, stxowner);
diff --git a/src/backend/commands/tablecmds.c b/src/backend/commands/tablecmds.c
index 7b6c69b7a5..73696a4210 100644
--- a/src/backend/commands/tablecmds.c
+++ b/src/backend/commands/tablecmds.c
@@ -3437,6 +3437,7 @@ StoreCatalogInheritance1(Oid relationId, Oid parentOid,
childobject.objectId = relationId;
childobject.objectSubId = 0;
+ depLockAndCheckObject(&parentobject);
recordDependencyOn(&childobject, &parentobject,
child_dependency_type(child_is_partition));
@@ -7440,6 +7441,7 @@ add_column_datatype_dependency(Oid relid, int32 attnum, Oid typid)
referenced.classId = TypeRelationId;
referenced.objectId = typid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
@@ -7461,6 +7463,7 @@ add_column_collation_dependency(Oid relid, int32 attnum, Oid collid)
referenced.classId = CollationRelationId;
referenced.objectId = collid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
}
@@ -10140,6 +10143,7 @@ addFkRecurseReferenced(List **wqueue, Constraint *fkconstraint, Relation rel,
ObjectAddress referenced;
ObjectAddressSet(referenced, ConstraintRelationId, parentConstr);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&address, &referenced, DEPENDENCY_INTERNAL);
}
@@ -10431,8 +10435,10 @@ addFkRecurseReferencing(List **wqueue, Constraint *fkconstraint, Relation rel,
*/
ObjectAddressSet(address, ConstraintRelationId, constrOid);
ObjectAddressSet(referenced, ConstraintRelationId, parentConstr);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&address, &referenced, DEPENDENCY_PARTITION_PRI);
ObjectAddressSet(referenced, RelationRelationId, partitionId);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&address, &referenced, DEPENDENCY_PARTITION_SEC);
/* Make all this visible before recursing */
@@ -10933,9 +10939,11 @@ CloneFkReferencing(List **wqueue, Relation parentRel, Relation partRel)
/* Set up partition dependencies for the new constraint */
ObjectAddressSet(address, ConstraintRelationId, constrOid);
ObjectAddressSet(referenced, ConstraintRelationId, parentConstrOid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&address, &referenced, DEPENDENCY_PARTITION_PRI);
ObjectAddressSet(referenced, RelationRelationId,
RelationGetRelid(partRel));
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&address, &referenced, DEPENDENCY_PARTITION_SEC);
/* Done with the cloned constraint's tuple */
@@ -14782,6 +14790,7 @@ ATExecSetAccessMethodNoStorage(Relation rel, Oid newAccessMethodId)
*/
ObjectAddressSet(relobj, RelationRelationId, reloid);
ObjectAddressSet(referenced, AccessMethodRelationId, rd_rel->relam);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&relobj, &referenced, DEPENDENCY_NORMAL);
}
else if (OidIsValid(oldAccessMethodId) &&
@@ -16400,6 +16409,7 @@ ATExecAddOf(Relation rel, const TypeName *ofTypename, LOCKMODE lockmode)
typeobj.classId = TypeRelationId;
typeobj.objectId = typeid;
typeobj.objectSubId = 0;
+ depLockAndCheckObject(&typeobj);
recordDependencyOn(&tableobj, &typeobj, DEPENDENCY_NORMAL);
/* Update pg_class.reloftype */
diff --git a/src/backend/commands/trigger.c b/src/backend/commands/trigger.c
index 95de402fa6..c0e7ddff1c 100644
--- a/src/backend/commands/trigger.c
+++ b/src/backend/commands/trigger.c
@@ -1018,8 +1018,6 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
((Form_pg_class) GETSTRUCT(tuple))->relhastriggers = true;
CatalogTupleUpdate(pgrel, &tuple->t_self, tuple);
-
- CommandCounterIncrement();
}
else
CacheInvalidateRelcacheByTuple(tuple);
@@ -1027,6 +1025,12 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
heap_freetuple(tuple);
table_close(pgrel, RowExclusiveLock);
+ /*
+ * CommandCounterIncrement here to ensure the new trigger entry is visible
+ * when we'll check of object existence when recording the dependencies.
+ */
+ CommandCounterIncrement();
+
/*
* If we're replacing a trigger, flush all the old dependencies before
* recording new ones.
@@ -1045,6 +1049,7 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
referenced.classId = ProcedureRelationId;
referenced.objectId = funcoid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
if (isInternal && OidIsValid(constraintOid))
@@ -1058,6 +1063,7 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
referenced.classId = ConstraintRelationId;
referenced.objectId = constraintOid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_INTERNAL);
}
else
@@ -1070,6 +1076,7 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
referenced.classId = RelationRelationId;
referenced.objectId = RelationGetRelid(rel);
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
if (OidIsValid(constrrelid))
@@ -1077,6 +1084,7 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
referenced.classId = RelationRelationId;
referenced.objectId = constrrelid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_AUTO);
}
/* Not possible to have an index dependency in this case */
@@ -1091,6 +1099,7 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
referenced.classId = ConstraintRelationId;
referenced.objectId = constraintOid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&myself);
recordDependencyOn(&referenced, &myself, DEPENDENCY_INTERNAL);
}
@@ -1100,8 +1109,10 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
if (OidIsValid(parentTriggerOid))
{
ObjectAddressSet(referenced, TriggerRelationId, parentTriggerOid);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_PARTITION_PRI);
ObjectAddressSet(referenced, RelationRelationId, RelationGetRelid(rel));
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_PARTITION_SEC);
}
}
@@ -1116,6 +1127,7 @@ CreateTriggerFiringOn(CreateTrigStmt *stmt, const char *queryString,
for (i = 0; i < ncolumns; i++)
{
referenced.objectSubId = columns[i];
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_NORMAL);
}
}
@@ -1255,9 +1267,11 @@ TriggerSetParentTrigger(Relation trigRel,
ObjectAddressSet(depender, TriggerRelationId, childTrigId);
ObjectAddressSet(referenced, TriggerRelationId, parentTrigId);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&depender, &referenced, DEPENDENCY_PARTITION_PRI);
ObjectAddressSet(referenced, RelationRelationId, childTableId);
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&depender, &referenced, DEPENDENCY_PARTITION_SEC);
}
else
diff --git a/src/backend/commands/tsearchcmds.c b/src/backend/commands/tsearchcmds.c
index b7b5019f1e..a62c997725 100644
--- a/src/backend/commands/tsearchcmds.c
+++ b/src/backend/commands/tsearchcmds.c
@@ -171,7 +171,7 @@ makeParserDependencies(HeapTuple tuple)
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
return myself;
@@ -329,7 +329,7 @@ makeDictionaryDependencies(HeapTuple tuple)
ObjectAddressSet(referenced, TSTemplateRelationId, dict->dicttemplate);
add_exact_object_address(&referenced, addrs);
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
return myself;
@@ -677,7 +677,7 @@ makeTSTemplateDependencies(HeapTuple tuple)
add_exact_object_address(&referenced, addrs);
}
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
return myself;
@@ -885,7 +885,7 @@ makeConfigurationDependencies(HeapTuple tuple, bool removeOld,
}
/* Record 'em (this includes duplicate elimination) */
- record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
+ lock_record_object_address_dependencies(&myself, addrs, DEPENDENCY_NORMAL);
free_object_addresses(addrs);
diff --git a/src/backend/commands/typecmds.c b/src/backend/commands/typecmds.c
index 2a1e713335..cbb61c1207 100644
--- a/src/backend/commands/typecmds.c
+++ b/src/backend/commands/typecmds.c
@@ -1794,6 +1794,7 @@ makeRangeConstructors(const char *name, Oid namespace,
* that they go away silently when the type is dropped. Note that
* pg_dump depends on this choice to avoid dumping the constructors.
*/
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_INTERNAL);
}
}
@@ -1859,6 +1860,7 @@ makeMultirangeConstructors(const char *name, Oid namespace,
* that they go away silently when the type is dropped. Note that pg_dump
* depends on this choice to avoid dumping the constructors.
*/
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_INTERNAL);
pfree(argtypes);
@@ -1898,6 +1900,7 @@ makeMultirangeConstructors(const char *name, Oid namespace,
1.0, /* procost */
0.0); /* prorows */
/* ditto */
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_INTERNAL);
pfree(argtypes);
*castFuncOid = myself.objectId;
@@ -1936,6 +1939,7 @@ makeMultirangeConstructors(const char *name, Oid namespace,
1.0, /* procost */
0.0); /* prorows */
/* ditto */
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced, DEPENDENCY_INTERNAL);
pfree(argtypes);
pfree(allParameterTypes);
diff --git a/src/backend/rewrite/rewriteDefine.c b/src/backend/rewrite/rewriteDefine.c
index 6cc9a8d8bf..53308dceb6 100644
--- a/src/backend/rewrite/rewriteDefine.c
+++ b/src/backend/rewrite/rewriteDefine.c
@@ -155,6 +155,7 @@ InsertRule(const char *rulname,
referenced.objectId = eventrel_oid;
referenced.objectSubId = 0;
+ depLockAndCheckObject(&referenced);
recordDependencyOn(&myself, &referenced,
(evtype == CMD_SELECT) ? DEPENDENCY_INTERNAL : DEPENDENCY_AUTO);
diff --git a/src/backend/utils/errcodes.txt b/src/backend/utils/errcodes.txt
index 3250d539e1..60e8539fe3 100644
--- a/src/backend/utils/errcodes.txt
+++ b/src/backend/utils/errcodes.txt
@@ -271,6 +271,7 @@ Section: Class 28 - Invalid Authorization Specification
Section: Class 2B - Dependent Privilege Descriptors Still Exist
2B000 E ERRCODE_DEPENDENT_PRIVILEGE_DESCRIPTORS_STILL_EXIST dependent_privilege_descriptors_still_exist
+2BP02 E ERRCODE_DEPENDENT_OBJECTS_DOES_NOT_EXIST dependent_objects_does_not_exist
2BP01 E ERRCODE_DEPENDENT_OBJECTS_STILL_EXIST dependent_objects_still_exist
Section: Class 2D - Invalid Transaction Termination
diff --git a/src/include/catalog/dependency.h b/src/include/catalog/dependency.h
index 7eee66f810..f404a92fb4 100644
--- a/src/include/catalog/dependency.h
+++ b/src/include/catalog/dependency.h
@@ -101,6 +101,8 @@ typedef struct ObjectAddresses ObjectAddresses;
/* in dependency.c */
extern void AcquireDeletionLock(const ObjectAddress *object, int flags);
+extern void depLockAndCheckObject(const ObjectAddress *object);
+extern void depLockAndCheckObjects(const ObjectAddress *object, int nobject);
extern void ReleaseDeletionLock(const ObjectAddress *object);
@@ -128,9 +130,9 @@ extern void add_exact_object_address(const ObjectAddress *object,
extern bool object_address_present(const ObjectAddress *object,
const ObjectAddresses *addrs);
-extern void record_object_address_dependencies(const ObjectAddress *depender,
- ObjectAddresses *referenced,
- DependencyType behavior);
+extern void lock_record_object_address_dependencies(const ObjectAddress *depender,
+ ObjectAddresses *referenced,
+ DependencyType behavior);
extern void sort_object_addresses(ObjectAddresses *addrs);
@@ -172,6 +174,7 @@ extern long changeDependenciesOf(Oid classId, Oid oldObjectId,
extern long changeDependenciesOn(Oid refClassId, Oid oldRefObjectId,
Oid newRefObjectId);
+extern bool isObjectPinned(const ObjectAddress *object);
extern Oid getExtensionOfObject(Oid classId, Oid objectId);
extern List *getAutoExtensionsOfObject(Oid classId, Oid objectId);
diff --git a/src/include/catalog/objectaddress.h b/src/include/catalog/objectaddress.h
index 3a70d80e32..56f746264b 100644
--- a/src/include/catalog/objectaddress.h
+++ b/src/include/catalog/objectaddress.h
@@ -53,6 +53,7 @@ extern void check_object_ownership(Oid roleid,
Node *object, Relation relation);
extern Oid get_object_namespace(const ObjectAddress *address);
+extern bool ObjectByIdExist(const ObjectAddress *address);
extern bool is_objectclass_supported(Oid class_id);
extern const char *get_object_class_descr(Oid class_id);
diff --git a/src/test/isolation/expected/test_dependencies_locks.out b/src/test/isolation/expected/test_dependencies_locks.out
new file mode 100644
index 0000000000..9b645d7aa5
--- /dev/null
+++ b/src/test/isolation/expected/test_dependencies_locks.out
@@ -0,0 +1,129 @@
+Parsed test spec with 2 sessions
+
+starting permutation: s1_begin s1_create_function_in_schema s2_drop_schema s1_commit
+step s1_begin: BEGIN;
+step s1_create_function_in_schema: CREATE FUNCTION testschema.foo() RETURNS int AS 'select 1' LANGUAGE sql;
+step s2_drop_schema: DROP SCHEMA testschema; <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_schema: <... completed>
+ERROR: cannot drop schema testschema because other objects depend on it
+
+starting permutation: s2_begin s2_drop_schema s1_create_function_in_schema s2_commit
+step s2_begin: BEGIN;
+step s2_drop_schema: DROP SCHEMA testschema;
+step s1_create_function_in_schema: CREATE FUNCTION testschema.foo() RETURNS int AS 'select 1' LANGUAGE sql; <waiting ...>
+step s2_commit: COMMIT;
+step s1_create_function_in_schema: <... completed>
+ERROR: schema testschema does not exist
+
+starting permutation: s1_begin s1_alter_function_schema s2_drop_alterschema s1_commit
+step s1_begin: BEGIN;
+step s1_alter_function_schema: ALTER FUNCTION public.falter() SET SCHEMA alterschema;
+step s2_drop_alterschema: DROP SCHEMA alterschema; <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_alterschema: <... completed>
+ERROR: cannot drop schema alterschema because other objects depend on it
+
+starting permutation: s2_begin s2_drop_alterschema s1_alter_function_schema s2_commit
+step s2_begin: BEGIN;
+step s2_drop_alterschema: DROP SCHEMA alterschema;
+step s1_alter_function_schema: ALTER FUNCTION public.falter() SET SCHEMA alterschema; <waiting ...>
+step s2_commit: COMMIT;
+step s1_alter_function_schema: <... completed>
+ERROR: schema alterschema does not exist
+
+starting permutation: s1_begin s1_create_function_with_argtype s2_drop_foo_type s1_commit
+step s1_begin: BEGIN;
+step s1_create_function_with_argtype: CREATE FUNCTION fooargtype(num foo) RETURNS int AS 'select 1' LANGUAGE sql;
+step s2_drop_foo_type: DROP TYPE public.foo; <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_foo_type: <... completed>
+ERROR: cannot drop type foo because other objects depend on it
+
+starting permutation: s2_begin s2_drop_foo_type s1_create_function_with_argtype s2_commit
+step s2_begin: BEGIN;
+step s2_drop_foo_type: DROP TYPE public.foo;
+step s1_create_function_with_argtype: CREATE FUNCTION fooargtype(num foo) RETURNS int AS 'select 1' LANGUAGE sql; <waiting ...>
+step s2_commit: COMMIT;
+step s1_create_function_with_argtype: <... completed>
+ERROR: type foo does not exist
+
+starting permutation: s1_begin s1_create_function_with_rettype s2_drop_foo_rettype s1_commit
+step s1_begin: BEGIN;
+step s1_create_function_with_rettype: CREATE FUNCTION footrettype() RETURNS id LANGUAGE sql RETURN 1;
+step s2_drop_foo_rettype: DROP DOMAIN id; <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_foo_rettype: <... completed>
+ERROR: cannot drop type id because other objects depend on it
+
+starting permutation: s2_begin s2_drop_foo_rettype s1_create_function_with_rettype s2_commit
+step s2_begin: BEGIN;
+step s2_drop_foo_rettype: DROP DOMAIN id;
+step s1_create_function_with_rettype: CREATE FUNCTION footrettype() RETURNS id LANGUAGE sql RETURN 1; <waiting ...>
+step s2_commit: COMMIT;
+step s1_create_function_with_rettype: <... completed>
+ERROR: type id does not exist
+
+starting permutation: s1_begin s1_create_function_with_function s2_drop_function_f s1_commit
+step s1_begin: BEGIN;
+step s1_create_function_with_function: CREATE FUNCTION foofunc() RETURNS int LANGUAGE SQL RETURN f() + 1;
+step s2_drop_function_f: DROP FUNCTION f(); <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_function_f: <... completed>
+ERROR: cannot drop function f() because other objects depend on it
+
+starting permutation: s2_begin s2_drop_function_f s1_create_function_with_function s2_commit
+step s2_begin: BEGIN;
+step s2_drop_function_f: DROP FUNCTION f();
+step s1_create_function_with_function: CREATE FUNCTION foofunc() RETURNS int LANGUAGE SQL RETURN f() + 1; <waiting ...>
+step s2_commit: COMMIT;
+step s1_create_function_with_function: <... completed>
+ERROR: function f() does not exist
+
+starting permutation: s1_begin s1_create_domain_with_domain s2_drop_domain_id s1_commit
+step s1_begin: BEGIN;
+step s1_create_domain_with_domain: CREATE DOMAIN idid as id;
+step s2_drop_domain_id: DROP DOMAIN id; <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_domain_id: <... completed>
+ERROR: cannot drop type id because other objects depend on it
+
+starting permutation: s2_begin s2_drop_domain_id s1_create_domain_with_domain s2_commit
+step s2_begin: BEGIN;
+step s2_drop_domain_id: DROP DOMAIN id;
+step s1_create_domain_with_domain: CREATE DOMAIN idid as id; <waiting ...>
+step s2_commit: COMMIT;
+step s1_create_domain_with_domain: <... completed>
+ERROR: type id does not exist
+
+starting permutation: s1_begin s1_create_table_with_type s2_drop_footab_type s1_commit
+step s1_begin: BEGIN;
+step s1_create_table_with_type: CREATE TABLE tabtype(a footab);
+step s2_drop_footab_type: DROP TYPE public.footab; <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_footab_type: <... completed>
+ERROR: cannot drop type footab because other objects depend on it
+
+starting permutation: s2_begin s2_drop_footab_type s1_create_table_with_type s2_commit
+step s2_begin: BEGIN;
+step s2_drop_footab_type: DROP TYPE public.footab;
+step s1_create_table_with_type: CREATE TABLE tabtype(a footab); <waiting ...>
+step s2_commit: COMMIT;
+step s1_create_table_with_type: <... completed>
+ERROR: type footab does not exist
+
+starting permutation: s1_begin s1_create_server_with_fdw_wrapper s2_drop_fdw_wrapper s1_commit
+step s1_begin: BEGIN;
+step s1_create_server_with_fdw_wrapper: CREATE SERVER srv_fdw_wrapper FOREIGN DATA WRAPPER fdw_wrapper;
+step s2_drop_fdw_wrapper: DROP FOREIGN DATA WRAPPER fdw_wrapper RESTRICT; <waiting ...>
+step s1_commit: COMMIT;
+step s2_drop_fdw_wrapper: <... completed>
+ERROR: cannot drop foreign-data wrapper fdw_wrapper because other objects depend on it
+
+starting permutation: s2_begin s2_drop_fdw_wrapper s1_create_server_with_fdw_wrapper s2_commit
+step s2_begin: BEGIN;
+step s2_drop_fdw_wrapper: DROP FOREIGN DATA WRAPPER fdw_wrapper RESTRICT;
+step s1_create_server_with_fdw_wrapper: CREATE SERVER srv_fdw_wrapper FOREIGN DATA WRAPPER fdw_wrapper; <waiting ...>
+step s2_commit: COMMIT;
+step s1_create_server_with_fdw_wrapper: <... completed>
+ERROR: foreign-data wrapper fdw_wrapper does not exist
diff --git a/src/test/isolation/isolation_schedule b/src/test/isolation/isolation_schedule
index 0342eb39e4..1b67f0bffe 100644
--- a/src/test/isolation/isolation_schedule
+++ b/src/test/isolation/isolation_schedule
@@ -114,3 +114,4 @@ test: serializable-parallel-2
test: serializable-parallel-3
test: matview-write-skew
test: lock-nowait
+test: test_dependencies_locks
diff --git a/src/test/isolation/specs/test_dependencies_locks.spec b/src/test/isolation/specs/test_dependencies_locks.spec
new file mode 100644
index 0000000000..5d04dfe9dc
--- /dev/null
+++ b/src/test/isolation/specs/test_dependencies_locks.spec
@@ -0,0 +1,89 @@
+setup
+{
+ CREATE SCHEMA testschema;
+ CREATE SCHEMA alterschema;
+ CREATE TYPE public.foo as enum ('one', 'two');
+ CREATE TYPE public.footab as enum ('three', 'four');
+ CREATE DOMAIN id AS int;
+ CREATE FUNCTION f() RETURNS int LANGUAGE SQL RETURN 1;
+ CREATE FUNCTION public.falter() RETURNS int LANGUAGE SQL RETURN 1;
+ CREATE FOREIGN DATA WRAPPER fdw_wrapper;
+}
+
+teardown
+{
+ DROP FUNCTION IF EXISTS testschema.foo();
+ DROP FUNCTION IF EXISTS fooargtype(num foo);
+ DROP FUNCTION IF EXISTS footrettype();
+ DROP FUNCTION IF EXISTS foofunc();
+ DROP FUNCTION IF EXISTS public.falter();
+ DROP FUNCTION IF EXISTS alterschema.falter();
+ DROP DOMAIN IF EXISTS idid;
+ DROP SERVER IF EXISTS srv_fdw_wrapper;
+ DROP TABLE IF EXISTS tabtype;
+ DROP SCHEMA IF EXISTS testschema;
+ DROP SCHEMA IF EXISTS alterschema;
+ DROP TYPE IF EXISTS public.foo;
+ DROP TYPE IF EXISTS public.footab;
+ DROP DOMAIN IF EXISTS id;
+ DROP FUNCTION IF EXISTS f();
+ DROP FOREIGN DATA WRAPPER IF EXISTS fdw_wrapper;
+}
+
+session "s1"
+
+step "s1_begin" { BEGIN; }
+step "s1_create_function_in_schema" { CREATE FUNCTION testschema.foo() RETURNS int AS 'select 1' LANGUAGE sql; }
+step "s1_create_function_with_argtype" { CREATE FUNCTION fooargtype(num foo) RETURNS int AS 'select 1' LANGUAGE sql; }
+step "s1_create_function_with_rettype" { CREATE FUNCTION footrettype() RETURNS id LANGUAGE sql RETURN 1; }
+step "s1_create_function_with_function" { CREATE FUNCTION foofunc() RETURNS int LANGUAGE SQL RETURN f() + 1; }
+step "s1_alter_function_schema" { ALTER FUNCTION public.falter() SET SCHEMA alterschema; }
+step "s1_create_domain_with_domain" { CREATE DOMAIN idid as id; }
+step "s1_create_table_with_type" { CREATE TABLE tabtype(a footab); }
+step "s1_create_server_with_fdw_wrapper" { CREATE SERVER srv_fdw_wrapper FOREIGN DATA WRAPPER fdw_wrapper; }
+step "s1_commit" { COMMIT; }
+
+session "s2"
+
+step "s2_begin" { BEGIN; }
+step "s2_drop_schema" { DROP SCHEMA testschema; }
+step "s2_drop_alterschema" { DROP SCHEMA alterschema; }
+step "s2_drop_foo_type" { DROP TYPE public.foo; }
+step "s2_drop_foo_rettype" { DROP DOMAIN id; }
+step "s2_drop_footab_type" { DROP TYPE public.footab; }
+step "s2_drop_function_f" { DROP FUNCTION f(); }
+step "s2_drop_domain_id" { DROP DOMAIN id; }
+step "s2_drop_fdw_wrapper" { DROP FOREIGN DATA WRAPPER fdw_wrapper RESTRICT; }
+step "s2_commit" { COMMIT; }
+
+# function - schema
+permutation "s1_begin" "s1_create_function_in_schema" "s2_drop_schema" "s1_commit"
+permutation "s2_begin" "s2_drop_schema" "s1_create_function_in_schema" "s2_commit"
+
+# alter function - schema
+permutation "s1_begin" "s1_alter_function_schema" "s2_drop_alterschema" "s1_commit"
+permutation "s2_begin" "s2_drop_alterschema" "s1_alter_function_schema" "s2_commit"
+
+# function - argtype
+permutation "s1_begin" "s1_create_function_with_argtype" "s2_drop_foo_type" "s1_commit"
+permutation "s2_begin" "s2_drop_foo_type" "s1_create_function_with_argtype" "s2_commit"
+
+# function - rettype
+permutation "s1_begin" "s1_create_function_with_rettype" "s2_drop_foo_rettype" "s1_commit"
+permutation "s2_begin" "s2_drop_foo_rettype" "s1_create_function_with_rettype" "s2_commit"
+
+# function - function
+permutation "s1_begin" "s1_create_function_with_function" "s2_drop_function_f" "s1_commit"
+permutation "s2_begin" "s2_drop_function_f" "s1_create_function_with_function" "s2_commit"
+
+# domain - domain
+permutation "s1_begin" "s1_create_domain_with_domain" "s2_drop_domain_id" "s1_commit"
+permutation "s2_begin" "s2_drop_domain_id" "s1_create_domain_with_domain" "s2_commit"
+
+# table - type
+permutation "s1_begin" "s1_create_table_with_type" "s2_drop_footab_type" "s1_commit"
+permutation "s2_begin" "s2_drop_footab_type" "s1_create_table_with_type" "s2_commit"
+
+# server - foreign data wrapper
+permutation "s1_begin" "s1_create_server_with_fdw_wrapper" "s2_drop_fdw_wrapper" "s1_commit"
+permutation "s2_begin" "s2_drop_fdw_wrapper" "s1_create_server_with_fdw_wrapper" "s2_commit"
--
2.34.1
--xlZXO4tPcOr9Z9/W--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v4] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 35b0b72f0a7..25ebcd3cc47 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 0e624fe36b9..677137207e7 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index d88a7a70308..153fbb477bb 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..d4c2f80eb46 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} PGLockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1;
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ PGLockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(PGLockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ PGLockFile *lock_file = (PGLockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index ffb413ab612..ad34142e7d6 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1956,6 +1956,7 @@ PGIOAlignedBlock
PGLZ_HistEntry
PGLZ_Strategy
PGLoadBalanceType
+PGLockFile
PGMessageField
PGModuleMagicFunction
PGNoticeHooks
base-commit: 73dfe79fd6034b1e7e41e83d9c82c166dba8eb67
--
2.52.0
--tks2nr37zts5e6h7--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
* [PATCH v3] Use open file description locks for lockfiles
@ 2025-12-18 17:21 Dmitrii Dolgov <[email protected]>
0 siblings, 0 replies; 286+ messages in thread
From: Dmitrii Dolgov @ 2025-12-18 17:21 UTC (permalink / raw)
When starting up, postmaster checks for an existing data directory lockfile. If
this file contains current process PID, it's assumed to be stale. Turns out
there is another possibility: we might be running in a PID namespace, and there
is another postgres running inside another PID namespace using the same data
directory. The result is that we don't see another process due to namespace
isolation and start concurrently with the other.
To prevent such situations, at startup use fcntl to get an exclusive open file
description lock for data directory lockfile. Since such locks are associated
with open file descriptors, meaning they're not affected by PID namespace
isolation. It's a "best effort" locking, intended to work with already existing
mechanism, not replace it.
This approach was discussed multiple times in the past, and usually was
rejected as the main work horse for the data directory lockfile due to:
* Portability issues. Open file description lock was a non-POSIX extension in
Linux and similar flock is from BSD standard. But looks like everybody agrees
that such locks make more sense than a typical advisory locks, and
F_OFD_SETLK made its way into POSIX.1 2024 [1].
* Issues with NFS. The current state of things here looks like this:
- NFSv3 doesn't implement open file description locks, they're converted to
advisory locks instead. Advisory locks are subject to namespace isolation,
meaning that processes in different PID namespaces will not see each other
advisory lock, and it's still possible to run multiple postgres
instances on the same data directory.
- NFSv4 uses a lease system for locking, I haven't found any mention of
conversion to advisory locks neither in the man page nor in RFC [2].
To summarize, the approach is now considered POSIX and should fix the described
problem everywhere, except NFSv3.
Use open file description lock for both data directory and socker
lockfiles, since both are affected in the same way.
[1]: https://pubs.opengroup.org/onlinepubs/9799919799/functions/fcntl.html
[2]: https://www.rfc-editor.org/rfc/rfc7530
Reviewed-by: Ilmar Yunusov <[email protected]>
---
configure | 14 +++
configure.ac | 3 +
meson.build | 1 +
src/backend/utils/init/miscinit.c | 136 ++++++++++++++++++++++++------
src/include/pg_config.h.in | 4 +
src/tools/pgindent/typedefs.list | 1 +
6 files changed, 134 insertions(+), 25 deletions(-)
diff --git a/configure b/configure
index 5f77f3cac29..15bda6c6413 100755
--- a/configure
+++ b/configure
@@ -16444,6 +16444,20 @@ cat >>confdefs.h <<_ACEOF
_ACEOF
+# Linux open file descriptor locks
+ac_fn_c_check_decl "$LINENO" "F_OFD_SETLK" "ac_cv_have_decl_F_OFD_SETLK" "#include <fcntl.h>
+"
+if test "x$ac_cv_have_decl_F_OFD_SETLK" = xyes; then :
+ ac_have_decl=1
+else
+ ac_have_decl=0
+fi
+
+cat >>confdefs.h <<_ACEOF
+#define HAVE_DECL_F_OFD_SETLK $ac_have_decl
+_ACEOF
+
+
ac_fn_c_check_func "$LINENO" "explicit_bzero" "ac_cv_func_explicit_bzero"
if test "x$ac_cv_func_explicit_bzero" = xyes; then :
$as_echo "#define HAVE_EXPLICIT_BZERO 1" >>confdefs.h
diff --git a/configure.ac b/configure.ac
index 61cee42daa7..e24cb7a6f01 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1913,6 +1913,9 @@ AC_CHECK_DECLS([memset_s], [], [], [#define __STDC_WANT_LIB_EXT1__ 1
# This is probably only present on macOS, but may as well check always
AC_CHECK_DECLS(F_FULLFSYNC, [], [], [#include <fcntl.h>])
+# Linux open file descriptor locks
+AC_CHECK_DECLS([F_OFD_SETLK], [], [], [#include <fcntl.h>])
+
AC_REPLACE_FUNCS(m4_normalize([
explicit_bzero
getopt
diff --git a/meson.build b/meson.build
index 568e0e150bf..3d641dc0403 100644
--- a/meson.build
+++ b/meson.build
@@ -2901,6 +2901,7 @@ decl_checks = [
['strlcpy', 'string.h'],
['strsep', 'string.h'],
['timingsafe_bcmp', 'string.h'],
+ ['F_OFD_SETLK', 'fcntl.h'],
]
# Need to check for function declarations for these functions, because
diff --git a/src/backend/utils/init/miscinit.c b/src/backend/utils/init/miscinit.c
index 7ffc808073a..26c3324542c 100644
--- a/src/backend/utils/init/miscinit.c
+++ b/src/backend/utils/init/miscinit.c
@@ -69,6 +69,15 @@ static List *lock_files = NIL;
static Latch LocalLatchData;
+typedef struct
+{
+ /* LockFile name. */
+ const char *filename;
+
+ /* File descriptor for open file description lock. */
+ int fd;
+} LockFile;
+
/* ----------------------------------------------------------------
* ignoring system indexes support stuff
*
@@ -1119,6 +1128,48 @@ RestoreClientConnectionInfo(char *conninfo)
*-------------------------------------------------------------------------
*/
+/*
+ * OFD lock the specified lockfile.
+ *
+ * Lock the lockfile with an open file description lock. If the lock is already
+ * taken, it's a hard stop. It's only a best effort test, and any other errors
+ * are ignored. On succes the file descriptor is duplicated, to make sure there
+ * will be at least one open copy of it to keep the lock.
+ *
+ * filename argument is used only for reporting purposes.
+ */
+static int
+OFDLockFile(int fd, const char *filename)
+{
+#if HAVE_DECL_F_OFD_SETLK
+ struct flock lock;
+
+ lock.l_type = F_WRLCK;
+ lock.l_whence = SEEK_SET;
+ lock.l_start = 0;
+ lock.l_len = 0;
+ lock.l_pid = 0;
+
+ if (fcntl(fd, F_OFD_SETLK, &lock) == -1)
+ {
+ if (errno == EAGAIN)
+ ereport(FATAL,
+ (errcode(ERRCODE_LOCK_FILE_EXISTS),
+ errmsg("cannot lock the lock file \"%s\"", filename),
+ errhint("Another server is starting.")));
+ else
+ {
+ elog(WARNING, "Failed locking file \"%s\", %m", filename);
+ return -1;
+ }
+ }
+ else
+ return dup(fd);
+#else
+ return -1
+#endif
+}
+
/*
* proc_exit callback to remove lockfiles.
*/
@@ -1129,9 +1180,16 @@ UnlinkLockFiles(int status, Datum arg)
foreach(l, lock_files)
{
- char *curfile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
- unlink(curfile);
+ /*
+ * Close the file descriptor, which keeps the open file description
+ * lock.
+ */
+ if (lock_file->fd > 0)
+ close(lock_file->fd);
+
+ unlink(lock_file->filename);
/* Should we complain if the unlink fails? */
}
/* Since we're about to exit, no need to reclaim storage */
@@ -1161,7 +1219,9 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *socketDir,
bool isDDLock, const char *refName)
{
- int fd;
+ int fd,
+ flock_fd = -1;
+ LockFile *lock_file;
char buffer[MAXPGPATH * 2 + 256];
int ntries;
int len;
@@ -1173,22 +1233,32 @@ CreateLockFile(const char *filename, bool amPostmaster,
const char *envvar;
/*
- * If the PID in the lockfile is our own PID or our parent's or
- * grandparent's PID, then the file must be stale (probably left over from
- * a previous system boot cycle). We need to check this because of the
- * likelihood that a reboot will assign exactly the same PID as we had in
- * the previous reboot, or one that's only one or two counts larger and
- * hence the lockfile's PID now refers to an ancestor shell process. We
- * allow pg_ctl to pass down its parent shell PID (our grandparent PID)
- * via the environment variable PG_GRANDPARENT_PID; this is so that
- * launching the postmaster via pg_ctl can be just as reliable as
- * launching it directly. There is no provision for detecting
- * further-removed ancestor processes, but if the init script is written
- * carefully then all but the immediate parent shell will be root-owned
- * processes and so the kill test will fail with EPERM. Note that we
- * cannot get a false negative this way, because an existing postmaster
- * would surely never launch a competing postmaster or pg_ctl process
- * directly.
+ * If we find an already existing lockfile containing our own PID, there
+ * are few options:
+ *
+ * - There is another process, that we don't see due to PID namespace
+ * isolation, which is already running in this data directory.
+ *
+ * To prevent two concurrent processes working with the same data
+ * directory, we first try to lock the lockfile exclusively.
+ *
+ * - The file must be stale, probably left over from a previous system
+ * boot cycle. The same if the lockfile contains our parent's or
+ * grandparent's PID.
+ *
+ * We need to check this because of the likelihood that a reboot will
+ * assign exactly the same PID as we had in the previous reboot, or one
+ * that's only one or two counts larger and hence the lockfile's PID now
+ * refers to an ancestor shell process. We allow pg_ctl to pass down its
+ * parent shell PID (our grandparent PID) via the environment variable
+ * PG_GRANDPARENT_PID; this is so that launching the postmaster via pg_ctl
+ * can be just as reliable as launching it directly. There is no
+ * provision for detecting further-removed ancestor processes, but if the
+ * init script is written carefully then all but the immediate parent
+ * shell will be root-owned processes and so the kill test will fail with
+ * EPERM. Note that we cannot get a false negative this way, because an
+ * existing postmaster would surely never launch a competing postmaster or
+ * pg_ctl process directly.
*/
my_pid = getpid();
@@ -1224,7 +1294,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
*/
fd = open(filename, O_RDWR | O_CREAT | O_EXCL, pg_file_create_mode);
if (fd >= 0)
- break; /* Success; exit the retry loop */
+ {
+ /* Success; lock and exit the retry loop */
+ flock_fd = OFDLockFile(fd, filename);
+ break;
+ }
/*
* Couldn't create the pid file. Probably it already exists.
@@ -1238,8 +1312,12 @@ CreateLockFile(const char *filename, bool amPostmaster,
/*
* Read the file to get the old owner's PID. Note race condition
* here: file might have been deleted since we tried to create it.
+ *
+ * We're going to use the same fd for flock, and want to create a
+ * write lock for the latter one. Since both fd and the lock have to
+ * be of the same type, open the file for read and write.
*/
- fd = open(filename, O_RDONLY, pg_file_create_mode);
+ fd = open(filename, O_RDWR, pg_file_create_mode);
if (fd < 0)
{
if (errno == ENOENT)
@@ -1249,6 +1327,10 @@ CreateLockFile(const char *filename, bool amPostmaster,
errmsg("could not open lock file \"%s\": %m",
filename)));
}
+
+ /* Try to lock the file. We stop here, if it's already locked. */
+ flock_fd = OFDLockFile(fd, filename);
+
pgstat_report_wait_start(WAIT_EVENT_LOCK_FILE_CREATE_READ);
if ((len = read(fd, buffer, sizeof(buffer) - 1)) < 0)
ereport(FATAL,
@@ -1448,7 +1530,11 @@ CreateLockFile(const char *filename, bool amPostmaster,
* Use lcons so that the lock files are unlinked in reverse order of
* creation; this is critical!
*/
- lock_files = lcons(pstrdup(filename), lock_files);
+ lock_file = palloc0_object(LockFile);
+ lock_file->filename = pstrdup(filename);
+ lock_file->fd = flock_fd;
+
+ lock_files = lcons(lock_file, lock_files);
}
/*
@@ -1495,14 +1581,14 @@ TouchSocketLockFiles(void)
foreach(l, lock_files)
{
- char *socketLockFile = (char *) lfirst(l);
+ LockFile *lock_file = (LockFile *) lfirst(l);
/* No need to touch the data directory lock file, we trust */
- if (strcmp(socketLockFile, DIRECTORY_LOCK_FILE) == 0)
+ if (strcmp(lock_file->filename, DIRECTORY_LOCK_FILE) == 0)
continue;
/* we just ignore any error here */
- (void) utime(socketLockFile, NULL);
+ (void) utime(lock_file->filename, NULL);
}
}
diff --git a/src/include/pg_config.h.in b/src/include/pg_config.h.in
index 4f8113c144b..cc38c06dc13 100644
--- a/src/include/pg_config.h.in
+++ b/src/include/pg_config.h.in
@@ -85,6 +85,10 @@
don't. */
#undef HAVE_DECL_F_FULLFSYNC
+/* Define to 1 if you have the declaration of `F_OFD_SETLK', and to 0 if you
+ don't. */
+#undef HAVE_DECL_F_OFD_SETLK
+
/* Define to 1 if you have the declaration of `memset_s', and to 0 if you
don't. */
#undef HAVE_DECL_MEMSET_S
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 1969d467c1d..185d69b5520 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -1653,6 +1653,7 @@ LocationLen
LockAcquireResult
LockClauseStrength
LockData
+LockFile
LockInfoData
LockInstanceData
LockMethod
base-commit: 031904048aa22e7c70dc8e9c170e2743f9b0f090
--
2.52.0
--zdel3ow7bygx53fm--
^ permalink raw reply [nested|flat] 286+ messages in thread
end of thread, other threads:[~2025-12-18 17:21 UTC | newest]
Thread overview: 286+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2024-03-29 15:43 [PATCH v9] Avoid orphaned objects dependencies Bertrand Drouvot <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v3] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
2025-12-18 17:21 [PATCH v4] Use open file description locks for lockfiles Dmitrii Dolgov <[email protected]>
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox