agora inbox for [email protected]  

help / color / mirror / Atom feed

[PATCH 1/2] Demonstrate possible race conditions in logical decoding.
305+ messages / 1 participants
[nested] [flat]

* [PATCH 1/2] Demonstrate possible race conditions in logical decoding.
@ 2026-01-19 11:54  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ messages in thread

From: Antonin Houska @ 2026-01-19 11:54 UTC (permalink / raw)

The problem is that the snapshot builder can create a snapshot before CLOG has
been updated. That breaks visibility check that use such snapshot. For more
details, see startup_race.spec.

Success of the test means that the problem is present. Thus it would need to
be modified if it should be merged into the tree.

Another problem that currently prevents this test from being merged is that it
hard-wires the logical decoding setup into the SET TRANSACTION command. I
tried to modify the isolation tester so it can use the logical replication
protocol (in which case the test could use the "CREATE_REPLICATION_SNAPSHOT
... (SNAPSHOT 'use')" command), but the tester does things that are not
compatible with that protocol (e.g. it sets the application_name parameter).
---
 .../test_decoding/expected/startup_race.out   |  85 ++++++++++++
 contrib/test_decoding/specs/startup_race.spec | 126 ++++++++++++++++++
 src/backend/access/transam/xact.c             |   6 +
 src/backend/replication/logical/snapbuild.c   |   3 +
 src/backend/utils/time/snapmgr.c              |  66 +++++++++
 src/include/utils/snapmgr.h                   |   1 +
 6 files changed, 287 insertions(+)
 create mode 100644 contrib/test_decoding/expected/startup_race.out
 create mode 100644 contrib/test_decoding/specs/startup_race.spec

diff --git a/contrib/test_decoding/expected/startup_race.out b/contrib/test_decoding/expected/startup_race.out
new file mode 100644
index 00000000000..597fa617831
--- /dev/null
+++ b/contrib/test_decoding/expected/startup_race.out
@@ -0,0 +1,85 @@
+Parsed test spec with 6 sessions
+
+starting permutation: s1_assign_xid s2_set_snapshot s3_assign_xid s1_rollback s3_rollback s4_do_changes s5_wake_up_full_snapshot s2_scan s2_rollback s5_wake_up_before_clog s6_check
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+injection_points_attach
+-----------------------
+                       
+(1 row)
+
+step s1_assign_xid: 
+	BEGIN;
+	CREATE TABLE b(i int);
+
+step s2_set_snapshot: 
+	BEGIN READ ONLY ISOLATION LEVEL REPEATABLE READ;
+	SET TRANSACTION SNAPSHOT 'from_slot';
+ <waiting ...>
+step s3_assign_xid: 
+	BEGIN;
+	CREATE TABLE c(i int);
+
+step s1_rollback: 
+	ROLLBACK;
+
+step s3_rollback: 
+	ROLLBACK;
+
+step s4_do_changes: 
+	INSERT INTO a(i, j) VALUES (3, 3);
+	UPDATE a SET j = j + 1 WHERE i = 1;
+	DELETE FROM a WHERE i = 2;
+ <waiting ...>
+step s5_wake_up_full_snapshot: 
+	SELECT injection_points_wakeup('snapbuild-full-snapshot');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step s2_set_snapshot: <... completed>
+step s2_scan: 
+	TABLE a;
+
+i|j
+-+-
+1|1
+2|2
+(2 rows)
+
+step s2_rollback: 
+	ROLLBACK;
+
+step s5_wake_up_before_clog: 
+	SELECT injection_points_wakeup('before-clog-update');
+
+injection_points_wakeup
+-----------------------
+                       
+(1 row)
+
+step s4_do_changes: <... completed>
+step s6_check: 
+	SELECT * FROM a ORDER BY i;
+
+i|j
+-+-
+1|1
+2|2
+(2 rows)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
+injection_points_detach
+-----------------------
+                       
+(1 row)
+
diff --git a/contrib/test_decoding/specs/startup_race.spec b/contrib/test_decoding/specs/startup_race.spec
new file mode 100644
index 00000000000..8f67e07fa7a
--- /dev/null
+++ b/contrib/test_decoding/specs/startup_race.spec
@@ -0,0 +1,126 @@
+setup
+{
+	CREATE TABLE a(i int primary key, j int) WITH (autovacuum_enabled = off);
+	INSERT INTO a(i, j) VALUES (1, 1), (2, 2);
+	CREATE EXTENSION injection_points;
+}
+
+session s1
+step s1_assign_xid
+{
+	BEGIN;
+	CREATE TABLE b(i int);
+}
+step s1_rollback
+{
+	ROLLBACK;
+}
+
+session s2
+setup
+{
+	SELECT injection_points_set_local();
+	SELECT injection_points_attach('snapbuild-full-snapshot', 'wait');
+}
+# Use special, hard-wired snapshot name to set the initial snapshot from
+# logical replication slot.
+step s2_set_snapshot
+{
+	BEGIN READ ONLY ISOLATION LEVEL REPEATABLE READ;
+	SET TRANSACTION SNAPSHOT 'from_slot';
+}
+# Perform the scan.
+step s2_scan
+{
+	TABLE a;
+}
+step s2_rollback
+{
+	ROLLBACK;
+}
+teardown
+{
+	SELECT injection_points_detach('snapbuild-full-snapshot');
+}
+
+session s3
+step s3_assign_xid
+{
+	BEGIN;
+	CREATE TABLE c(i int);
+}
+step s3_rollback
+{
+	ROLLBACK;
+}
+
+session s4
+setup
+{
+	SELECT injection_points_set_local();
+	SELECT injection_points_attach('before-clog-update', 'wait');
+}
+step s4_do_changes
+{
+	INSERT INTO a(i, j) VALUES (3, 3);
+	UPDATE a SET j = j + 1 WHERE i = 1;
+	DELETE FROM a WHERE i = 2;
+}
+teardown
+{
+	SELECT injection_points_detach('before-clog-update');
+}
+
+session s5
+step s5_wake_up_full_snapshot
+{
+	SELECT injection_points_wakeup('snapbuild-full-snapshot');
+}
+step s5_wake_up_before_clog
+{
+	SELECT injection_points_wakeup('before-clog-update');
+}
+
+session s6
+step s6_check
+{
+	SELECT * FROM a ORDER BY i;
+}
+
+permutation
+# Let the snapshot builder go through all the states. The problematic case
+# happens in the FULL_SNAPSHOT.
+s1_assign_xid
+# This should leave the builder in BUILDING_SNAPSHOT, waiting for the active
+# transaction to end.
+s2_set_snapshot
+# Make sure that s1_rollback does not allow going to CONSISTENT directly.
+s3_assign_xid
+# Let the builder proceed to FULL_SNAPSHOT. It should stop at the injection
+# point 'snapbuild-full-snapshot'.
+s1_rollback
+# The transaction of s3 is not needed anymore, CONSISTENT should be the next
+# stage.
+s3_rollback
+# Perform data changes before the snapshot builder triggers creation of the
+# RUNNING_XACTS record. This will stop before setting transaction status in
+# CLOG.
+s4_do_changes
+# Unblock the injection point so that the snapshot can finally be created.
+s5_wake_up_full_snapshot
+# Use the snapshot for a scan. The snapshot will consider s4 not running
+# anymore, however CLOG is not aware of the commit yet. Thus
+# HeapTupleSatisfiesMVCC will consider the transaction aborted. In particular,
+# for UPDATE, if both xmax of the old version and xmin of the new version are
+# considered aborted, so the effects of the UPDATE are lost
+# altogether. Similarly, INSERT and DELETE have no effect because the xmin
+# transaction of the inserted tuple and xmax of the deleted tuple are
+# considered aborted
+s2_scan
+s2_rollback
+# CLOG can be updated now.
+s5_wake_up_before_clog
+# Scan the table again using a new transaction, with a normal transaction
+# snapshot. The results are still wrong due to hint bits set incorrectly.
+s6_check
+
diff --git a/src/backend/access/transam/xact.c b/src/backend/access/transam/xact.c
index c857e23552f..2fea45b2fed 100644
--- a/src/backend/access/transam/xact.c
+++ b/src/backend/access/transam/xact.c
@@ -65,6 +65,7 @@
 #include "utils/builtins.h"
 #include "utils/combocid.h"
 #include "utils/guc.h"
+#include "utils/injection_point.h"
 #include "utils/inval.h"
 #include "utils/memutils.h"
 #include "utils/relmapper.h"
@@ -1348,6 +1349,9 @@ RecordTransactionCommit(void)
 													 &RelcacheInitFileInval);
 	wrote_xlog = (XactLastRecEnd != 0);
 
+	/* Load the injection point before entering the critical section */
+	INJECTION_POINT_LOAD("before-clog-update");
+
 	/*
 	 * If we haven't been assigned an XID yet, we neither can, nor do we want
 	 * to write a COMMIT record.
@@ -1514,6 +1518,8 @@ RecordTransactionCommit(void)
 	{
 		XLogFlush(XactLastRecEnd);
 
+		INJECTION_POINT_CACHED("before-clog-update", NULL);
+
 		/*
 		 * Now we may update the CLOG, if we wrote a COMMIT record above
 		 */
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index 7f79621b57e..9b09dc8eac1 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -141,6 +141,7 @@
 #include "storage/procarray.h"
 #include "storage/standby.h"
 #include "utils/builtins.h"
+#include "utils/injection_point.h"
 #include "utils/memutils.h"
 #include "utils/snapmgr.h"
 #include "utils/snapshot.h"
@@ -1387,6 +1388,8 @@ SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *runn
 				errdetail("Waiting for transactions (approximately %d) older than %u to end.",
 						  running->xcnt, running->nextXid));
 
+		INJECTION_POINT("snapbuild-full-snapshot", NULL);
+
 		SnapBuildWaitSnapshot(running, running->nextXid);
 	}
 
diff --git a/src/backend/utils/time/snapmgr.c b/src/backend/utils/time/snapmgr.c
index 2e6197f5f35..f327b779004 100644
--- a/src/backend/utils/time/snapmgr.c
+++ b/src/backend/utils/time/snapmgr.c
@@ -110,10 +110,13 @@
 #include "access/subtrans.h"
 #include "access/transam.h"
 #include "access/xact.h"
+#include "access/xlogutils.h"
 #include "datatype/timestamp.h"
 #include "lib/pairingheap.h"
 #include "miscadmin.h"
 #include "port/pg_lfind.h"
+#include "replication/logical.h"
+#include "replication/snapbuild.h"
 #include "storage/fd.h"
 #include "storage/predicate.h"
 #include "storage/proc.h"
@@ -1421,6 +1424,17 @@ ImportSnapshot(const char *idstr)
 				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 				 errmsg("a snapshot-importing transaction must have isolation level SERIALIZABLE or REPEATABLE READ")));
 
+	if (strcmp(idstr, "from_slot") == 0)
+	{
+		Snapshot	snap;
+
+		snap = create_test_snapshot();
+		/* XXX sourcevxid shouldn't be needed in this special case */
+		SetTransactionSnapshot(snap, NULL, MyProcPid, MyProc);
+
+		return;
+	}
+
 	/*
 	 * Verify the identifier: only 0-9, A-F and hyphens are allowed.  We do
 	 * this mainly to prevent reading arbitrary files.
@@ -1969,3 +1983,55 @@ ResOwnerReleaseSnapshot(Datum res)
 {
 	UnregisterSnapshotNoOwner((Snapshot) DatumGetPointer(res));
 }
+
+/*
+ * CreateReplicationSlot() with the CRS_USE_SNAPSHOT option would be useful
+ * for testing, but regression tests cannot speak both replication and query
+ * protocol at the same time. This function can be used instead to create a
+ * snapshot for special tests of logical replication.
+ */
+Snapshot
+create_test_snapshot(void)
+{
+	const	char	*slotname = "test_slot";
+	const	char *plugin;
+	LogicalDecodingContext *ctx;
+	Snapshot	snap;
+
+	/*
+	 * XXX Hard-wired values are fine for the special test that needs this
+	 * function.
+	 */
+	plugin = "test_decoding";
+
+	Assert(!MyReplicationSlot);
+
+	CheckLogicalDecodingRequirements();
+
+	ReplicationSlotCreate(slotname, true, RS_TEMPORARY,
+						  false, false, false);
+
+	/*
+	 * Ensure the logical decoding is enabled before initializing the
+	 * logical decoding context.
+	 */
+	EnsureLogicalDecodingEnabled();
+	Assert(IsLogicalDecodingEnabled());
+
+	ctx = CreateInitDecodingContext(plugin, NIL, true,
+									InvalidXLogRecPtr,
+									XL_ROUTINE(.page_read = read_local_xlog_page,
+											   .segment_open = wal_segment_open,
+											   .segment_close = wal_segment_close),
+									NULL, NULL, NULL);
+
+	/* build initial snapshot, might take a while */
+	DecodingContextFindStartpoint(ctx);
+
+	/* Do what the function is called for. */
+	snap = SnapBuildInitialSnapshot(ctx->snapshot_builder);
+	snap = CopySnapshot(snap);
+	FreeDecodingContext(ctx);
+
+	return snap;
+}
diff --git a/src/include/utils/snapmgr.h b/src/include/utils/snapmgr.h
index b8c01a291a1..9f0d60ebc1b 100644
--- a/src/include/utils/snapmgr.h
+++ b/src/include/utils/snapmgr.h
@@ -123,4 +123,5 @@ extern Snapshot RestoreSnapshot(char *start_address);
 struct PGPROC;
 extern void RestoreTransactionSnapshot(Snapshot snapshot, struct PGPROC *source_pgproc);
 
+extern Snapshot create_test_snapshot(void);
 #endif							/* SNAPMGR_H */
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread

* [PATCH] Teach REPACK to upgrade its lock safely.
@ 2026-04-14 09:59  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 305+ 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] 305+ messages in thread