agora inbox for pgsql-hackers@postgresql.org  

help / color / mirror / Atom feed

[PATCH] Add pg_tablespace_avail() functions
377+ messages / 2 participants
[nested] [flat]

* [PATCH] Add pg_tablespace_avail() functions
@ 2025-03-14 15:29 Christoph Berg <myon@debian.org>
  0 siblings, 0 replies; 377+ messages in thread

From: Christoph Berg @ 2025-03-14 15:29 UTC (permalink / raw)

This exposes the f_avail value from statvfs() on tablespace directories
on the SQL level, allowing monitoring of free disk space from within the
server. On windows, GetDiskFreeSpaceEx() is used.

Permissions required match those from pg_tablespace_size().

In psql, include a new "Free" column in \db+ output.

Add test coverage for pg_tablespace_avail() and the previously not
covered pg_tablespace_size() function.
---
 doc/src/sgml/func.sgml                   | 21 ++++++
 doc/src/sgml/ref/psql-ref.sgml           |  2 +-
 src/backend/utils/adt/dbsize.c           | 94 ++++++++++++++++++++++++
 src/bin/psql/describe.c                  | 11 ++-
 src/include/catalog/pg_proc.dat          |  8 ++
 src/test/regress/expected/tablespace.out | 21 ++++++
 src/test/regress/sql/tablespace.sql      | 10 +++
 7 files changed, 163 insertions(+), 4 deletions(-)

diff --git a/doc/src/sgml/func.sgml b/doc/src/sgml/func.sgml
index 51dd8ad6571..0b4456ad958 100644
--- a/doc/src/sgml/func.sgml
+++ b/doc/src/sgml/func.sgml
@@ -30093,6 +30093,27 @@ postgres=# SELECT '0/0'::pg_lsn + pd.segment_number * ps.setting::int + :offset
        </para></entry>
       </row>
 
+      <row>
+       <entry role="func_table_entry"><para role="func_signature">
+        <indexterm>
+         <primary>pg_tablespace_avail</primary>
+        </indexterm>
+        <function>pg_tablespace_avail</function> ( <type>name</type> )
+        <returnvalue>bigint</returnvalue>
+       </para>
+       <para role="func_signature">
+        <function>pg_tablespace_avail</function> ( <type>oid</type> )
+        <returnvalue>bigint</returnvalue>
+       </para>
+       <para>
+        Returns the available disk space in the tablespace with the
+        specified name or OID. To use this function, you must
+        have <literal>CREATE</literal> privilege on the specified tablespace
+        or have privileges of the <literal>pg_read_all_stats</literal> role,
+        unless it is the default tablespace for the current database.
+       </para></entry>
+      </row>
+
       <row>
        <entry role="func_table_entry"><para role="func_signature">
         <indexterm>
diff --git a/doc/src/sgml/ref/psql-ref.sgml b/doc/src/sgml/ref/psql-ref.sgml
index cedccc14129..9e1bec0b422 100644
--- a/doc/src/sgml/ref/psql-ref.sgml
+++ b/doc/src/sgml/ref/psql-ref.sgml
@@ -1492,7 +1492,7 @@ SELECT $1 \parse stmt1
         If <literal>x</literal> is appended to the command name, the results
         are displayed in expanded mode.
         If <literal>+</literal> is appended to the command name, each tablespace
-        is listed with its associated options, on-disk size, permissions and
+        is listed with its associated options, on-disk size and free disk space, permissions and
         description.
         </para>
         </listitem>
diff --git a/src/backend/utils/adt/dbsize.c b/src/backend/utils/adt/dbsize.c
index 25865b660ef..30e0cb8d111 100644
--- a/src/backend/utils/adt/dbsize.c
+++ b/src/backend/utils/adt/dbsize.c
@@ -12,6 +12,11 @@
 #include "postgres.h"
 
 #include <sys/stat.h>
+#ifdef WIN32
+#include <fileapi.h>
+#else
+#include <sys/statvfs.h>
+#endif
 
 #include "access/htup_details.h"
 #include "access/relation.h"
@@ -316,6 +321,95 @@ pg_tablespace_size_name(PG_FUNCTION_ARGS)
 }
 
 
+/*
+ * Return available disk space of tablespace. Returns -1 if the tablespace
+ * directory cannot be found.
+ */
+static int64
+calculate_tablespace_avail(Oid tblspcOid)
+{
+	char		tblspcPath[MAXPGPATH];
+	AclResult	aclresult;
+#ifdef WIN32
+	ULARGE_INTEGER lpFreeBytesAvailable;
+#else
+	struct statvfs fst;
+#endif
+
+	/*
+	 * User must have privileges of pg_read_all_stats or have CREATE privilege
+	 * for target tablespace, either explicitly granted or implicitly because
+	 * it is default for current database.
+	 */
+	if (tblspcOid != MyDatabaseTableSpace &&
+		!has_privs_of_role(GetUserId(), ROLE_PG_READ_ALL_STATS))
+	{
+		aclresult = object_aclcheck(TableSpaceRelationId, tblspcOid, GetUserId(), ACL_CREATE);
+		if (aclresult != ACLCHECK_OK)
+			aclcheck_error(aclresult, OBJECT_TABLESPACE,
+						   get_tablespace_name(tblspcOid));
+	}
+
+	if (tblspcOid == DEFAULTTABLESPACE_OID)
+		snprintf(tblspcPath, MAXPGPATH, "base");
+	else if (tblspcOid == GLOBALTABLESPACE_OID)
+		snprintf(tblspcPath, MAXPGPATH, "global");
+	else
+		snprintf(tblspcPath, MAXPGPATH, "%s/%u/%s", PG_TBLSPC_DIR, tblspcOid,
+				 TABLESPACE_VERSION_DIRECTORY);
+
+#ifdef WIN32
+	if (GetDiskFreeSpaceEx(tblspcPath, &lpFreeBytesAvailable, NULL, NULL) == false)
+		return -1;
+
+	return lpFreeBytesAvailable.QuadPart; /* ULONGLONG part of ULARGE_INTEGER */
+#else
+	if (statvfs(tblspcPath, &fst) < 0)
+		return -1;
+
+	return fst.f_bavail * fst.f_frsize; /* available blocks times fragment size */
+#endif
+}
+
+Datum
+pg_tablespace_avail_oid(PG_FUNCTION_ARGS)
+{
+	Oid			tblspcOid = PG_GETARG_OID(0);
+	int64		avail;
+
+	/*
+	 * Not needed for correctness, but avoid non-user-facing error message
+	 * later if the tablespace doesn't exist.
+	 */
+	if (!SearchSysCacheExists1(TABLESPACEOID, ObjectIdGetDatum(tblspcOid)))
+		ereport(ERROR,
+				errcode(ERRCODE_UNDEFINED_OBJECT),
+				errmsg("tablespace with OID %u does not exist", tblspcOid));
+
+	avail = calculate_tablespace_avail(tblspcOid);
+
+	if (avail < 0)
+		PG_RETURN_NULL();
+
+	PG_RETURN_INT64(avail);
+}
+
+Datum
+pg_tablespace_avail_name(PG_FUNCTION_ARGS)
+{
+	Name		tblspcName = PG_GETARG_NAME(0);
+	Oid			tblspcOid = get_tablespace_oid(NameStr(*tblspcName), false);
+	int64		avail;
+
+	avail = calculate_tablespace_avail(tblspcOid);
+
+	if (avail < 0)
+		PG_RETURN_NULL();
+
+	PG_RETURN_INT64(avail);
+}
+
+
 /*
  * calculate size of (one fork of) a relation
  *
diff --git a/src/bin/psql/describe.c b/src/bin/psql/describe.c
index e6cf468ac9e..8c52a126ac1 100644
--- a/src/bin/psql/describe.c
+++ b/src/bin/psql/describe.c
@@ -241,10 +241,15 @@ describeTablespaces(const char *pattern, bool verbose)
 		printACLColumn(&buf, "spcacl");
 		appendPQExpBuffer(&buf,
 						  ",\n  spcoptions AS \"%s\""
-						  ",\n  pg_catalog.pg_size_pretty(pg_catalog.pg_tablespace_size(oid)) AS \"%s\""
-						  ",\n  pg_catalog.shobj_description(oid, 'pg_tablespace') AS \"%s\"",
+						  ",\n  pg_catalog.pg_size_pretty(pg_catalog.pg_tablespace_size(oid)) AS \"%s\"",
 						  gettext_noop("Options"),
-						  gettext_noop("Size"),
+						  gettext_noop("Size"));
+		if (pset.sversion >= 180000)
+			appendPQExpBuffer(&buf,
+							  ",\n  pg_catalog.pg_size_pretty(pg_catalog.pg_tablespace_avail(oid)) AS \"%s\"",
+							  gettext_noop("Free"));
+		appendPQExpBuffer(&buf,
+						  ",\n  pg_catalog.shobj_description(oid, 'pg_tablespace') AS \"%s\"",
 						  gettext_noop("Description"));
 	}
 
diff --git a/src/include/catalog/pg_proc.dat b/src/include/catalog/pg_proc.dat
index 42e427f8fe8..9d64da6bfb8 100644
--- a/src/include/catalog/pg_proc.dat
+++ b/src/include/catalog/pg_proc.dat
@@ -7680,6 +7680,14 @@
   descr => 'total disk space usage for the specified tablespace',
   proname => 'pg_tablespace_size', provolatile => 'v', prorettype => 'int8',
   proargtypes => 'name', prosrc => 'pg_tablespace_size_name' },
+{ oid => '6015',
+  descr => 'disk stats for the specified tablespace',
+  proname => 'pg_tablespace_avail', provolatile => 'v', prorettype => 'int8',
+  proargtypes => 'oid', prosrc => 'pg_tablespace_avail_oid' },
+{ oid => '6016',
+  descr => 'disk stats for the specified tablespace',
+  proname => 'pg_tablespace_avail', provolatile => 'v', prorettype => 'int8',
+  proargtypes => 'name', prosrc => 'pg_tablespace_avail_name' },
 { oid => '2324', descr => 'total disk space usage for the specified database',
   proname => 'pg_database_size', provolatile => 'v', prorettype => 'int8',
   proargtypes => 'oid', prosrc => 'pg_database_size_oid' },
diff --git a/src/test/regress/expected/tablespace.out b/src/test/regress/expected/tablespace.out
index a90e39e5738..6709ed794df 100644
--- a/src/test/regress/expected/tablespace.out
+++ b/src/test/regress/expected/tablespace.out
@@ -20,6 +20,27 @@ SELECT spcoptions FROM pg_tablespace WHERE spcname = 'regress_tblspacewith';
  {random_page_cost=3.0}
 (1 row)
 
+-- check size functions
+SELECT pg_tablespace_size('pg_default') BETWEEN 1_000_000 and 10_000_000_000, -- rough sanity check
+       pg_tablespace_size('pg_global') BETWEEN 100_000 and 10_000_000,
+       pg_tablespace_size('regress_tblspacewith'); -- empty
+ ?column? | ?column? | pg_tablespace_size 
+----------+----------+--------------------
+ t        | t        |                  0
+(1 row)
+
+SELECT pg_tablespace_size('missing');
+ERROR:  tablespace "missing" does not exist
+SELECT pg_tablespace_avail('pg_default') > 1_000_000,
+       pg_tablespace_avail('pg_global') > 1_000_000,
+       pg_tablespace_avail('regress_tblspacewith') > 1_000_000;
+ ?column? | ?column? | ?column? 
+----------+----------+----------
+ t        | t        | t
+(1 row)
+
+SELECT pg_tablespace_avail('missing');
+ERROR:  tablespace "missing" does not exist
 -- drop the tablespace so we can re-use the location
 DROP TABLESPACE regress_tblspacewith;
 -- This returns a relative path as of an effect of allow_in_place_tablespaces,
diff --git a/src/test/regress/sql/tablespace.sql b/src/test/regress/sql/tablespace.sql
index dfe3db096e2..3fcd4bb00ff 100644
--- a/src/test/regress/sql/tablespace.sql
+++ b/src/test/regress/sql/tablespace.sql
@@ -17,6 +17,16 @@ CREATE TABLESPACE regress_tblspacewith LOCATION '' WITH (random_page_cost = 3.0)
 -- check to see the parameter was used
 SELECT spcoptions FROM pg_tablespace WHERE spcname = 'regress_tblspacewith';
 
+-- check size functions
+SELECT pg_tablespace_size('pg_default') BETWEEN 1_000_000 and 10_000_000_000, -- rough sanity check
+       pg_tablespace_size('pg_global') BETWEEN 100_000 and 10_000_000,
+       pg_tablespace_size('regress_tblspacewith'); -- empty
+SELECT pg_tablespace_size('missing');
+SELECT pg_tablespace_avail('pg_default') > 1_000_000,
+       pg_tablespace_avail('pg_global') > 1_000_000,
+       pg_tablespace_avail('regress_tblspacewith') > 1_000_000;
+SELECT pg_tablespace_avail('missing');
+
 -- drop the tablespace so we can re-use the location
 DROP TABLESPACE regress_tblspacewith;
 
-- 
2.47.2


--65NGEM4fhwMw73YU--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59 Antonin Houska <ah@cybertec.at>
  0 siblings, 0 replies; 377+ messages in thread

From: Antonin Houska @ 2026-04-14 09:59 UTC (permalink / raw)

REPACK (CONCURRENTLY) needs to upgrade its ShareUpdateExclusiveLock (SUEL) to
AccessExclusiveLock at the end of table processing. If another session, which
already has a non-conflicting lock on the table, tried to get SUEL too, it
would end up in a deadlock with REPACK. Such situation does not hurt as long
as the deadlock detector choses to terminate the other session, but it's
possible that it terminates REPACK. A lot of REPACK's work would be wasted
this way.

This patch checks for such situation, and if the other session tries to get
the SUEL, it receives a deadlock report before it actually starts waiting.

Please note that the other session can safely run VACUUM w/o receiving the
deadlock report. The point is that VACUUM cannot run in a transaction block,
so the session cannot have any other lock when trying to get SUEL on the
table. In that case, REPACK gets its AccessExclusiveLock first, even if VACUUM
requested SUEL earlier. VACUUM then just needs wait for REPACK to finish, and
then it can start.
---
 src/backend/commands/repack.c                 | 25 +++++++-
 src/backend/storage/lmgr/deadlock.c           | 10 ++--
 src/backend/storage/lmgr/lock.c               | 30 ++++++++++
 src/backend/storage/lmgr/proc.c               | 55 ++++++++++++++++-
 src/include/storage/lock.h                    |  5 +-
 src/include/storage/proc.h                    |  6 +-
 .../injection_points/expected/repack.out      | 59 ++++++++++++++++++-
 .../injection_points/specs/repack.spec        | 29 +++++++++
 8 files changed, 209 insertions(+), 10 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..e9193067666 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -285,6 +285,18 @@ ExecRepack(ParseState *pstate, RepackStmt *stmt, bool isTopLevel)
 		 * to understand and we don't lose any functionality.
 		 */
 		PreventInTransactionBlock(isTopLevel, "REPACK (CONCURRENTLY)");
+
+		/*
+		 * Also set the PROC_IN_CONCURRENT_REPACK flag.  This makes the lock
+		 * manager cause anyone that would conflict with us to error out.
+		 * It's important to set this flag ahead of actually locking the
+		 * relation; it won't of course affect anyone until we do have a lock
+		 * that others can conflict with.
+		 */
+		LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+		MyProc->statusFlags |= PROC_IN_CONCURRENT_REPACK;
+		ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+		LWLockRelease(ProcArrayLock);
 	}
 
 	/*
@@ -3086,7 +3098,18 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 		LockRelationOid(OldHeap->rd_rel->reltoastrelid, AccessExclusiveLock);
 
 	/*
-	 * Tuples and pages of the old heap will be gone, but the heap will stay.
+	 * Now that we have all access-exclusive locks on all relations, we no
+	 * longer want other processes to error out when trying to acquire a
+	 * conflicting lock.  Therefore, unset our flag.
+	 */
+	LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
+	MyProc->statusFlags &= ~PROC_IN_CONCURRENT_REPACK;
+	ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
+	LWLockRelease(ProcArrayLock);
+
+	/*
+	 * Tuples and pages of the old heap will be gone, but the heap itself will
+	 * stay.
 	 */
 	TransferPredicateLocksToHeapRelation(OldHeap);
 	foreach_ptr(RelationData, index, indexrels)
diff --git a/src/backend/storage/lmgr/deadlock.c b/src/backend/storage/lmgr/deadlock.c
index b8962d875b6..3772a7aa3df 100644
--- a/src/backend/storage/lmgr/deadlock.c
+++ b/src/backend/storage/lmgr/deadlock.c
@@ -1146,8 +1146,10 @@ DeadLockReport(void)
 void
 RememberSimpleDeadLock(PGPROC *proc1,
 					   LOCKMODE lockmode,
-					   LOCK *lock,
-					   PGPROC *proc2)
+					   LOCK *lock, /* XXX Only lockmode? */
+					   PGPROC *proc2,
+					   LOCKMODE lockmode2,
+					   LOCKTAG *locktag2)
 {
 	DEADLOCK_INFO *info = &deadlockDetails[0];
 
@@ -1155,8 +1157,8 @@ RememberSimpleDeadLock(PGPROC *proc1,
 	info->lockmode = lockmode;
 	info->pid = proc1->pid;
 	info++;
-	info->locktag = proc2->waitLock->tag;
-	info->lockmode = proc2->waitLockMode;
+	info->locktag = *locktag2;
+	info->lockmode = lockmode2;
 	info->pid = proc2->pid;
 	nDeadlockDetails = 2;
 }
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..44aa2ca26a7 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -674,6 +674,36 @@ LockHeldByMe(const LOCKTAG *locktag,
 	return false;
 }
 
+/*
+ * HaveRelLock -- test whether the current backend has a fast path lock on
+ *		the relation in any mode.
+ */
+bool
+HaveFastPathLock(Oid relid)
+{
+	uint32		group = FAST_PATH_REL_GROUP(relid);
+	bool		result = false;
+
+	LWLockAcquire(&MyProc->fpInfoLock, LW_SHARED);
+	for (int i = 0; i < FP_LOCK_SLOTS_PER_GROUP; i++)
+	{
+		uint32		f = FAST_PATH_SLOT(group, i);
+		uint32		lockmask;
+
+		if (relid != MyProc->fpRelId[f])
+			continue;
+		lockmask = FAST_PATH_GET_BITS(MyProc, f);
+		if (lockmask)
+		{
+			result = true;
+			break;
+		}
+	}
+	LWLockRelease(&MyProc->fpInfoLock);
+
+	return result;
+}
+
 #ifdef USE_ASSERT_CHECKING
 /*
  * GetLockMethodLocalHash -- return the hash of local locks, for modules that
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..3d47b1284a3 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -1156,6 +1156,7 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 	LOCKMASK	myHeldLocks;
 	bool		early_deadlock = false;
 	PGPROC	   *leader = MyProc->lockGroupLeader;
+	bool		check_repack = false;
 
 	Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
 
@@ -1193,6 +1194,56 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 		}
 	}
 
+	/*
+	 * If I am already holding this lock (in any mode) and trying to get
+	 * ShareUpdateExclusiveLock (or higher), check if REPACK (CONCURRENTLY) is
+	 * already holding ShareUpdateExclusiveLock. The problem is if that by
+	 * trying to upgrade the lock to AccessExclusiveLock it would get into a
+	 * deadlock with us.
+	 */
+	if (lock->tag.locktag_type == LOCKTAG_RELATION &&
+		lock->tag.locktag_field1 == MyDatabaseId &&
+		lockmode >= ShareUpdateExclusiveLock)
+	{
+		if (myHeldLocks > 0 ||
+			HaveFastPathLock(lock->tag.locktag_field2))
+			check_repack = true;
+	}
+	if (check_repack)
+	{
+		dlist_iter	iter;
+
+		dlist_foreach(iter, &lock->procLocks)
+		{
+			PROCLOCK   *otherproclock;
+			PGPROC		*otherproc;
+
+			otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
+			otherproc = otherproclock->tag.myProc;
+			if (otherproc != MyProc &&
+				otherproc->statusFlags & PROC_IN_CONCURRENT_REPACK)
+			{
+				LOCKMASK	repackmask = otherproclock->holdMask;
+
+				/*
+				 * Is REPACK already holding ShareUpdateExclusiveLock?
+				 */
+				if ((repackmask & LOCKBIT_ON(ShareUpdateExclusiveLock)) != 0)
+				{
+					/*
+					 * There should be no more than one REPACK working on
+					 * particular table, so let's error out.
+					 */
+					RememberSimpleDeadLock(MyProc, lockmode, lock,
+										   otherproc,
+										   ShareUpdateExclusiveLock,
+										   &otherproclock->tag.myLock->tag);
+					return PROC_WAIT_STATUS_ERROR;
+				}
+			}
+		}
+	}
+
 	/*
 	 * Determine where to add myself in the wait queue.
 	 *
@@ -1240,7 +1291,9 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 					 * a flag to check below, and break out of loop.  Also,
 					 * record deadlock info for later message.
 					 */
-					RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
+					RememberSimpleDeadLock(MyProc, lockmode, lock, proc,
+										   proc->waitLockMode,
+										   &proc->waitLock->tag);
 					early_deadlock = true;
 					break;
 				}
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..85804d61ce9 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -401,6 +401,7 @@ extern void LockReleaseCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern void LockReassignCurrentOwner(LOCALLOCK **locallocks, int nlocks);
 extern bool LockHeldByMe(const LOCKTAG *locktag,
 						 LOCKMODE lockmode, bool orstronger);
+extern bool HaveFastPathLock(Oid relid);
 #ifdef USE_ASSERT_CHECKING
 extern HTAB *GetLockMethodLocalHash(void);
 #endif
@@ -440,7 +441,9 @@ pg_noreturn extern void DeadLockReport(void);
 extern void RememberSimpleDeadLock(PGPROC *proc1,
 								   LOCKMODE lockmode,
 								   LOCK *lock,
-								   PGPROC *proc2);
+								   PGPROC *proc2,
+								   LOCKMODE lockmode2,
+								   LOCKTAG *locktag2);
 extern void InitDeadLockChecking(void);
 
 extern int	LockWaiterCount(const LOCKTAG *locktag);
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..76c6bb44251 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -70,10 +70,12 @@ struct XidCache
 #define		PROC_AFFECTS_ALL_HORIZONS	0x20	/* this proc's xmin must be
 												 * included in vacuum horizons
 												 * in all databases */
+#define		PROC_IN_CONCURRENT_REPACK	0x40	/* REPACK (CONCURRENTLY) */
 
-/* flags reset at EOXact */
+/* flags reset at EOXact.  A bit of a misnomer ... */
 #define		PROC_VACUUM_STATE_MASK \
-	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND)
+	(PROC_IN_VACUUM | PROC_IN_SAFE_IC | PROC_VACUUM_FOR_WRAPAROUND | \
+	 PROC_IN_CONCURRENT_REPACK)
 
 /*
  * Xmin-related flags. Make sure any flags that affect how the process' Xmin
diff --git a/src/test/modules/injection_points/expected/repack.out b/src/test/modules/injection_points/expected/repack.out
index b575e9052ee..960044776f4 100644
--- a/src/test/modules/injection_points/expected/repack.out
+++ b/src/test/modules/injection_points/expected/repack.out
@@ -1,4 +1,4 @@
-Parsed test spec with 2 sessions
+Parsed test spec with 3 sessions
 
 starting permutation: wait_before_lock change_existing change_new change_subxact1 change_subxact2 check2 wakeup_before_lock check1
 injection_points_attach
@@ -111,3 +111,60 @@ injection_points_detach
                        
 (1 row)
 
+
+starting permutation: check1_relnode_only wait_before_lock lock3 wakeup_before_lock check1_relnode_only
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    1
+(1 row)
+
+step wait_before_lock: 
+	REPACK (CONCURRENTLY) repack_test USING INDEX repack_test_pkey;
+ <waiting ...>
+step lock3: 
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+
+i|j
+-+-
+1|1
+(1 row)
+
+ERROR:  deadlock detected
+step wakeup_before_lock: 
+	SELECT injection_points_wakeup('repack-concurrently-before-lock');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step wait_before_lock: <... completed>
+step check1_relnode_only: 
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+
+count
+-----
+    2
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/src/test/modules/injection_points/specs/repack.spec b/src/test/modules/injection_points/specs/repack.spec
index d727a9b056b..5ffcb558718 100644
--- a/src/test/modules/injection_points/specs/repack.spec
+++ b/src/test/modules/injection_points/specs/repack.spec
@@ -55,6 +55,13 @@ step check1
 	FROM data_s1 d1 FULL JOIN data_s2 d2 USING (i, j)
 	WHERE d1.i ISNULL OR d2.i ISNULL;
 }
+step check1_relnode_only
+{
+	INSERT INTO relfilenodes(node)
+	SELECT relfilenode FROM pg_class WHERE relname='repack_test';
+
+	SELECT count(DISTINCT node) FROM relfilenodes;
+}
 teardown
 {
 	SELECT injection_points_detach('repack-concurrently-before-lock');
@@ -129,6 +136,20 @@ step wakeup_before_lock
 	SELECT injection_points_wakeup('repack-concurrently-before-lock');
 }
 
+session s3
+# Try to acquire SharedUpdateExclusiveLock on a table while REPACK is already
+# holding it and before it tries to upgrade it to AccessExclusiveLock. Since
+# this session is already holding another lock no the table, REPACK cannot
+# resolve the problem by getting ahead in the wait queue. The solution is that
+# the LOCK TABLE triggers a deadlock report. The deadlock does not actually
+# take place, but it would if this session started to wait.
+step lock3
+{
+	BEGIN;
+	SELECT * FROM repack_test ORDER BY i LIMIT 1;
+	LOCK TABLE repack_test IN SHARE UPDATE EXCLUSIVE MODE;
+}
+
 # Test if data changes introduced while one session is performing REPACK
 # CONCURRENTLY find their way into the table.
 permutation
@@ -140,3 +161,11 @@ permutation
 	check2
 	wakeup_before_lock
 	check1
+
+# See 'lock3' above. We check relnodes to make sure that REPACK finished.
+permutation
+	check1_relnode_only
+	wait_before_lock
+	lock3
+	wakeup_before_lock
+	check1_relnode_only
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 377+ messages in thread