agora inbox for [email protected]  

help / color / mirror / Atom feed

[PATCH v56 1/3] Allow logical replication snapshots to be database-specific
200+ messages / 2 participants
[nested] [flat]

* [PATCH v56 1/3] Allow logical replication snapshots to be database-specific
@ 2026-04-06 22:07  Álvaro Herrera <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Álvaro Herrera @ 2026-04-06 22:07 UTC (permalink / raw)

By default, the logical decoding assumes access to shared catalogs, so
the snapshot builder needs to consider cluster-wide XIDs during startup.
That in turn means that, if any transaction is already running (and has
XID assigned), the snapshot builder needs to wait for its completion, as
it does not know if that transaction performed catalog changes earlier.

A possible problem with this concept is that if REPACK (CONCURRENTLY) is
running in some database, backends running the same command in other
databases get stuck until the first one has committed. Thus only a
single backend in the cluster can run REPACK (CONCURRENTLY) at any time.
Likewise, REPACK (CONCURRENTLY) can block walsenders starting on behalf
of subscriptions throughout the cluster.

This patch adds a new option to logical replication output plugin, to
declare that it does not use shared catalogs (i.e. catalogs that can be
changed by transactions running in other databases in the cluster). In
that case, no snapshot the backend will use during the decoding needs to
contain information about transactions running in other databases. Thus
the snapshot builder only needs to wait for completion of transactions
in the current database.

Currently we only use this option in the REPACK background worker. It
could possibly be used in the plugin for logical replication too,
however that would need thorough analysis of that plugin.

Bump WAL version number, due to a new field in xl_running_xacts.

Author: Antonin Houska <[email protected]>
Reviewed-by: Álvaro Herrera <[email protected]>
Discussion: https://postgr.es/m/90475.1775218118@localhost
---
 contrib/pg_visibility/pg_visibility.c       |  4 +-
 doc/src/sgml/logicaldecoding.sgml           |  4 ++
 src/backend/access/index/genam.c            |  8 +++
 src/backend/access/rmgrdesc/standbydesc.c   |  2 +
 src/backend/access/transam/xlog.c           |  2 +-
 src/backend/access/transam/xlogfuncs.c      |  2 +-
 src/backend/postmaster/bgwriter.c           |  2 +-
 src/backend/replication/logical/decode.c    | 17 +++++-
 src/backend/replication/logical/logical.c   |  3 +
 src/backend/replication/logical/snapbuild.c | 63 ++++++++++++++++++++-
 src/backend/replication/pgrepack/pgrepack.c |  7 +++
 src/backend/replication/slot.c              |  2 +-
 src/backend/storage/ipc/procarray.c         | 23 +++++++-
 src/backend/storage/ipc/standby.c           | 24 +++++++-
 src/include/access/genam.h                  |  1 +
 src/include/access/xlog_internal.h          |  2 +-
 src/include/miscadmin.h                     |  9 +++
 src/include/replication/output_plugin.h     |  1 +
 src/include/replication/snapbuild.h         |  3 +-
 src/include/storage/procarray.h             |  2 +-
 src/include/storage/standby.h               |  3 +-
 src/include/storage/standbydefs.h           |  1 +
 22 files changed, 167 insertions(+), 18 deletions(-)

diff --git a/contrib/pg_visibility/pg_visibility.c b/contrib/pg_visibility/pg_visibility.c
index dfab0b64cf5..d564bd2a00c 100644
--- a/contrib/pg_visibility/pg_visibility.c
+++ b/contrib/pg_visibility/pg_visibility.c
@@ -621,7 +621,7 @@ GetStrictOldestNonRemovableTransactionId(Relation rel)
 	else if (rel == NULL || rel->rd_rel->relisshared)
 	{
 		/* Shared relation: take into account all running xids */
-		runningTransactions = GetRunningTransactionData();
+		runningTransactions = GetRunningTransactionData(InvalidOid);
 		LWLockRelease(ProcArrayLock);
 		LWLockRelease(XidGenLock);
 		return runningTransactions->oldestRunningXid;
@@ -632,7 +632,7 @@ GetStrictOldestNonRemovableTransactionId(Relation rel)
 		 * Normal relation: take into account xids running within the current
 		 * database
 		 */
-		runningTransactions = GetRunningTransactionData();
+		runningTransactions = GetRunningTransactionData(InvalidOid);
 		LWLockRelease(ProcArrayLock);
 		LWLockRelease(XidGenLock);
 		return runningTransactions->oldestDatabaseRunningXid;
diff --git a/doc/src/sgml/logicaldecoding.sgml b/doc/src/sgml/logicaldecoding.sgml
index 6dc49108997..9b1d68d0de6 100644
--- a/doc/src/sgml/logicaldecoding.sgml
+++ b/doc/src/sgml/logicaldecoding.sgml
@@ -959,6 +959,7 @@ typedef struct OutputPluginOptions
 {
     OutputPluginOutputType output_type;
     bool        receive_rewrites;
+    bool        need_shared_catalogs;
 } OutputPluginOptions;
 </programlisting>
       <literal>output_type</literal> has to either be set to
@@ -969,6 +970,9 @@ typedef struct OutputPluginOptions
       also be called for changes made by heap rewrites during certain DDL
       operations.  These are of interest to plugins that handle DDL
       replication, but they require special handling.
+      <literal>need_shared_catalogs</literal> can be set to false if you are
+      certain the plugin functions do not access shared system catalogs.
+      Doing so can speed up creation of replication slots that use this plugin.
      </para>
 
      <para>
diff --git a/src/backend/access/index/genam.c b/src/backend/access/index/genam.c
index 1408989c568..97d44b84622 100644
--- a/src/backend/access/index/genam.c
+++ b/src/backend/access/index/genam.c
@@ -394,6 +394,14 @@ systable_beginscan(Relation heapRelation,
 	SysScanDesc sysscan;
 	Relation	irel;
 
+	/*
+	 * If this backend promised that it won't access shared catalogs during
+	 * logical decoding, this it the right place to verify.
+	 */
+	Assert(!HistoricSnapshotActive() ||
+		   accessSharedCatalogsInDecoding ||
+		   !heapRelation->rd_rel->relisshared);
+
 	if (indexOK &&
 		!IgnoreSystemIndexes &&
 		!ReindexIsProcessingIndex(indexId))
diff --git a/src/backend/access/rmgrdesc/standbydesc.c b/src/backend/access/rmgrdesc/standbydesc.c
index 0a291354ae2..685d1bdb024 100644
--- a/src/backend/access/rmgrdesc/standbydesc.c
+++ b/src/backend/access/rmgrdesc/standbydesc.c
@@ -41,6 +41,8 @@ standby_desc_running_xacts(StringInfo buf, xl_running_xacts *xlrec)
 		for (i = 0; i < xlrec->subxcnt; i++)
 			appendStringInfo(buf, " %u", xlrec->xids[xlrec->xcnt + i]);
 	}
+
+	appendStringInfo(buf, "; dbid: %u", xlrec->dbid);
 }
 
 void
diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c
index b82af9a85c0..3f08a832ca6 100644
--- a/src/backend/access/transam/xlog.c
+++ b/src/backend/access/transam/xlog.c
@@ -7735,7 +7735,7 @@ CreateCheckPoint(int flags)
 	 * recovery we don't need to write running xact data.
 	 */
 	if (!shutdown && XLogStandbyInfoActive())
-		LogStandbySnapshot();
+		LogStandbySnapshot(InvalidOid);
 
 	START_CRIT_SECTION();
 
diff --git a/src/backend/access/transam/xlogfuncs.c b/src/backend/access/transam/xlogfuncs.c
index 65bbaeda59c..0f5979691e6 100644
--- a/src/backend/access/transam/xlogfuncs.c
+++ b/src/backend/access/transam/xlogfuncs.c
@@ -245,7 +245,7 @@ pg_log_standby_snapshot(PG_FUNCTION_ARGS)
 				(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 				 errmsg("pg_log_standby_snapshot() can only be used if \"wal_level\" >= \"replica\"")));
 
-	recptr = LogStandbySnapshot();
+	recptr = LogStandbySnapshot(InvalidOid);
 
 	/*
 	 * As a convenience, return the WAL location of the last inserted record
diff --git a/src/backend/postmaster/bgwriter.c b/src/backend/postmaster/bgwriter.c
index 1d8947774a9..a30de4262eb 100644
--- a/src/backend/postmaster/bgwriter.c
+++ b/src/backend/postmaster/bgwriter.c
@@ -289,7 +289,7 @@ BackgroundWriterMain(const void *startup_data, size_t startup_data_len)
 			if (now >= timeout &&
 				last_snapshot_lsn <= GetLastImportantRecPtr())
 			{
-				last_snapshot_lsn = LogStandbySnapshot();
+				last_snapshot_lsn = LogStandbySnapshot(InvalidOid);
 				last_snapshot_ts = now;
 			}
 		}
diff --git a/src/backend/replication/logical/decode.c b/src/backend/replication/logical/decode.c
index c9fea8cad28..38c5a4f5540 100644
--- a/src/backend/replication/logical/decode.c
+++ b/src/backend/replication/logical/decode.c
@@ -382,7 +382,16 @@ standby_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf)
 			{
 				xl_running_xacts *running = (xl_running_xacts *) XLogRecGetData(r);
 
-				SnapBuildProcessRunningXacts(builder, buf->origptr, running);
+				/*
+				 * Update this decoder's idea of transactions currently
+				 * running.  In doing so we will determine whether we have
+				 * reached consistent status.
+				 *
+				 * If the output plugin doesn't need access to shared
+				 * catalogs, we can ignore transactions in other databases.
+				 */
+				SnapBuildProcessRunningXacts(builder, buf->origptr, running,
+											 !ctx->options.need_shared_catalogs);
 
 				/*
 				 * Abort all transactions that we keep track of, that are
@@ -392,8 +401,12 @@ standby_decode(LogicalDecodingContext *ctx, XLogRecordBuffer *buf)
 				 * all running transactions which includes prepared ones,
 				 * while shutdown checkpoints just know that no non-prepared
 				 * transactions are in progress.
+				 *
+				 * The database-specific records might work here too, but it's
+				 * not their purpose.
 				 */
-				ReorderBufferAbortOld(ctx->reorder, running->oldestRunningXid);
+				if (!OidIsValid(running->dbid))
+					ReorderBufferAbortOld(ctx->reorder, running->oldestRunningXid);
 			}
 			break;
 		case XLOG_STANDBY_LOCK:
diff --git a/src/backend/replication/logical/logical.c b/src/backend/replication/logical/logical.c
index f20d0c542f3..8ceaf64d164 100644
--- a/src/backend/replication/logical/logical.c
+++ b/src/backend/replication/logical/logical.c
@@ -285,6 +285,9 @@ StartupDecodingContext(List *output_plugin_options,
 	ctx->write = do_write;
 	ctx->update_progress = update_progress;
 
+	/* Assume shared catalog access. The startup callback can change it. */
+	ctx->options.need_shared_catalogs = true;
+
 	ctx->output_plugin_options = output_plugin_options;
 
 	ctx->fast_forward = fast_forward;
diff --git a/src/backend/replication/logical/snapbuild.c b/src/backend/replication/logical/snapbuild.c
index b4269a3b102..ffb4ab2cf2a 100644
--- a/src/backend/replication/logical/snapbuild.c
+++ b/src/backend/replication/logical/snapbuild.c
@@ -154,6 +154,14 @@
 static ResourceOwner SavedResourceOwnerDuringExport = NULL;
 static bool ExportInProgress = false;
 
+/*
+ * If a backend is going to do logical decoding and the output plugin does
+ * not need to access shared catalogs, setting this variable to false can make
+ * the decoding startup faster. In particular, the backend will not need to
+ * wait for completion of already running transactions in other databases.
+ */
+bool		accessSharedCatalogsInDecoding = true;
+
 /* ->committed and ->catchange manipulation */
 static void SnapBuildPurgeOlderTxn(SnapBuild *builder);
 
@@ -170,7 +178,8 @@ static inline bool SnapBuildXidHasCatalogChanges(SnapBuild *builder, Transaction
 												 uint32 xinfo);
 
 /* xlog reading helper functions for SnapBuildProcessRunningXacts */
-static bool SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running);
+static bool SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn,
+								  xl_running_xacts *running);
 static void SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff);
 
 /* serialization functions */
@@ -226,6 +235,9 @@ AllocateSnapshotBuilder(ReorderBuffer *reorder,
 
 	MemoryContextSwitchTo(oldcontext);
 
+	/* The default is that shared catalog are used. */
+	accessSharedCatalogsInDecoding = true;
+
 	return builder;
 }
 
@@ -244,6 +256,9 @@ FreeSnapshotBuilder(SnapBuild *builder)
 		builder->snapshot = NULL;
 	}
 
+	/* The default is that shared catalog are used. */
+	accessSharedCatalogsInDecoding = true;
+
 	/* other resources are deallocated via memory context reset */
 	MemoryContextDelete(context);
 }
@@ -1136,7 +1151,8 @@ SnapBuildXidHasCatalogChanges(SnapBuild *builder, TransactionId xid,
  * anymore.
  */
 void
-SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running)
+SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running,
+							 bool db_specific)
 {
 	ReorderBufferTXN *txn;
 	TransactionId xmin;
@@ -1148,6 +1164,33 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
 	 */
 	if (builder->state < SNAPBUILD_CONSISTENT)
 	{
+		/*
+		 * To reduce the potential for unnecessarily waiting for completion of
+		 * unrelated transactions, the caller can declare that only
+		 * transactions of the current database are relevant at this stage.
+		 */
+		if (db_specific)
+		{
+			/*
+			 * If we must only keep track of transactions running in the
+			 * current database, we need transaction info from exactly that
+			 * database.
+			 */
+			if (running->dbid != MyDatabaseId)
+			{
+				LogStandbySnapshot(MyDatabaseId);
+
+				return;
+			}
+
+			/*
+			 * We'd better be able to check during scan if the plugin does not
+			 * lie.
+			 */
+			if (accessSharedCatalogsInDecoding)
+				accessSharedCatalogsInDecoding = false;
+		}
+
 		/* returns false if there's no point in performing cleanup just yet */
 		if (!SnapBuildFindSnapshot(builder, lsn, running))
 			return;
@@ -1155,6 +1198,16 @@ SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xact
 	else
 		SnapBuildSerialize(builder, lsn);
 
+	/*
+	 * Database specific transaction info may exist to reach CONSISTENT state
+	 * faster, however the code below makes no use of it. Moreover, such
+	 * record might cause problems because the following normal (cluster-wide)
+	 * record can have lower value of oldestRunningXid. In that case, let's
+	 * wait with the cleanup for the next regular cluster-wide record.
+	 */
+	if (OidIsValid(running->dbid))
+		return;
+
 	/*
 	 * Update range of interesting xids based on the running xacts
 	 * information. We don't increase ->xmax using it, because once we are in
@@ -1465,7 +1518,11 @@ SnapBuildWaitSnapshot(xl_running_xacts *running, TransactionId cutoff)
 	 */
 	if (!RecoveryInProgress())
 	{
-		LogStandbySnapshot();
+		/*
+		 * If the last transaction info was about specific database, so needs
+		 * to be the next one - at least until we're in the CONSISTENT state.
+		 */
+		LogStandbySnapshot(running->dbid);
 	}
 }
 
diff --git a/src/backend/replication/pgrepack/pgrepack.c b/src/backend/replication/pgrepack/pgrepack.c
index 457b6518a8e..4da66aa0f44 100644
--- a/src/backend/replication/pgrepack/pgrepack.c
+++ b/src/backend/replication/pgrepack/pgrepack.c
@@ -52,6 +52,13 @@ repack_startup(LogicalDecodingContext *ctx, OutputPluginOptions *opt,
 	/* Probably unnecessary, as we don't use the SQL interface ... */
 	opt->output_type = OUTPUT_PLUGIN_BINARY_OUTPUT;
 
+	/*
+	 * REPACK doesn't need access to shared catalogs, so we can speed up the
+	 * historic snapshot creation by setting this flag.  We'll only have to
+	 * wait for transactions in our database.
+	 */
+	opt->need_shared_catalogs = false;
+
 	if (ctx->output_plugin_options != NIL)
 	{
 		ereport(ERROR,
diff --git a/src/backend/replication/slot.c b/src/backend/replication/slot.c
index 21a213a0ebf..a1f37e59dbc 100644
--- a/src/backend/replication/slot.c
+++ b/src/backend/replication/slot.c
@@ -1756,7 +1756,7 @@ ReplicationSlotReserveWal(void)
 		XLogRecPtr	flushptr;
 
 		/* make sure we have enough information to start */
-		flushptr = LogStandbySnapshot();
+		flushptr = LogStandbySnapshot(InvalidOid);
 
 		/* and make sure it's fsynced to disk */
 		XLogFlush(flushptr);
diff --git a/src/backend/storage/ipc/procarray.c b/src/backend/storage/ipc/procarray.c
index f540bb6b23f..9299bcebbda 100644
--- a/src/backend/storage/ipc/procarray.c
+++ b/src/backend/storage/ipc/procarray.c
@@ -2623,9 +2623,11 @@ ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
  *
  * Note that if any transaction has overflowed its cached subtransactions
  * then there is no real need include any subtransactions.
+ *
+ * If 'dbid' is valid, only gather transactions running in that database.
  */
 RunningTransactions
-GetRunningTransactionData(void)
+GetRunningTransactionData(Oid dbid)
 {
 	/* result workspace */
 	static RunningTransactionsData CurrentRunningXactsData;
@@ -2700,6 +2702,18 @@ GetRunningTransactionData(void)
 		if (!TransactionIdIsValid(xid))
 			continue;
 
+		/*
+		 * Filter by database OID if requested.
+		 */
+		if (OidIsValid(dbid))
+		{
+			int			pgprocno = arrayP->pgprocnos[index];
+			PGPROC	   *proc = &allProcs[pgprocno];
+
+			if (proc->databaseId != dbid)
+				continue;
+		}
+
 		/*
 		 * Be careful not to exclude any xids before calculating the values of
 		 * oldestRunningXid and suboverflowed, since these are used to clean
@@ -2750,6 +2764,12 @@ GetRunningTransactionData(void)
 			PGPROC	   *proc = &allProcs[pgprocno];
 			int			nsubxids;
 
+			/*
+			 * Filter by database OID if requested.
+			 */
+			if (OidIsValid(dbid) && proc->databaseId != dbid)
+				continue;
+
 			/*
 			 * Save subtransaction XIDs. Other backends can't add or remove
 			 * entries while we're holding XidGenLock.
@@ -2783,6 +2803,7 @@ GetRunningTransactionData(void)
 	 * increases if slots do.
 	 */
 
+	CurrentRunningXacts->dbid = dbid;
 	CurrentRunningXacts->xcnt = count - subcount;
 	CurrentRunningXacts->subxcnt = subcount;
 	CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
diff --git a/src/backend/storage/ipc/standby.c b/src/backend/storage/ipc/standby.c
index de9092fdf5b..29af7733948 100644
--- a/src/backend/storage/ipc/standby.c
+++ b/src/backend/storage/ipc/standby.c
@@ -1188,6 +1188,14 @@ standby_redo(XLogReaderState *record)
 		xl_running_xacts *xlrec = (xl_running_xacts *) XLogRecGetData(record);
 		RunningTransactionsData running;
 
+		/*
+		 * Records issued for specific database are not suitable for physical
+		 * replication because that affects the whole cluster. In particular,
+		 * the list of XID is probably incomplete here.
+		 */
+		if (OidIsValid(xlrec->dbid))
+			return;
+
 		running.xcnt = xlrec->xcnt;
 		running.subxcnt = xlrec->subxcnt;
 		running.subxid_status = xlrec->subxid_overflow ? SUBXIDS_MISSING : SUBXIDS_IN_ARRAY;
@@ -1277,11 +1285,22 @@ standby_redo(XLogReaderState *record)
  * as there's no independent knob to just enable logical decoding. For
  * details of how this is used, check snapbuild.c's introductory comment.
  *
+ * If 'dbid' is valid, only gather transactions running in that
+ * database. snapbuild.c can use such running xacts information for faster
+ * startup, but it still needs normal (cluster-wide) during the actual
+ * decoding - see standby_decode() and SnapBuildProcessRunningXacts() for
+ * details. Other processes (e.g. checkpointer) issue the cluster-wide records
+ * whether logical decoding is active or not.
+ *
+ * Please be careful about using this argument for other purposes. In
+ * particular, physical replication *must* ignore the database-specific
+ * records, exactly because they do not cover the whole cluster - see
+ * standby_redo().
  *
  * Returns the RecPtr of the last inserted record.
  */
 XLogRecPtr
-LogStandbySnapshot(void)
+LogStandbySnapshot(Oid dbid)
 {
 	XLogRecPtr	recptr;
 	RunningTransactions running;
@@ -1314,7 +1333,7 @@ LogStandbySnapshot(void)
 	 * Log details of all in-progress transactions. This should be the last
 	 * record we write, because standby will open up when it sees this.
 	 */
-	running = GetRunningTransactionData();
+	running = GetRunningTransactionData(dbid);
 
 	/*
 	 * GetRunningTransactionData() acquired ProcArrayLock, we must release it.
@@ -1358,6 +1377,7 @@ LogCurrentRunningXacts(RunningTransactions CurrRunningXacts)
 	xl_running_xacts xlrec;
 	XLogRecPtr	recptr;
 
+	xlrec.dbid = CurrRunningXacts->dbid;
 	xlrec.xcnt = CurrRunningXacts->xcnt;
 	xlrec.subxcnt = CurrRunningXacts->subxcnt;
 	xlrec.subxid_overflow = (CurrRunningXacts->subxid_status != SUBXIDS_IN_ARRAY);
diff --git a/src/include/access/genam.h b/src/include/access/genam.h
index b69320a7fc8..de55821e414 100644
--- a/src/include/access/genam.h
+++ b/src/include/access/genam.h
@@ -136,6 +136,7 @@ typedef struct IndexOrderByDistance
 	bool		isnull;
 } IndexOrderByDistance;
 
+
 /*
  * generalized index_ interface routines (in indexam.c)
  */
diff --git a/src/include/access/xlog_internal.h b/src/include/access/xlog_internal.h
index 10c18d39ff8..13ae3ad4fbb 100644
--- a/src/include/access/xlog_internal.h
+++ b/src/include/access/xlog_internal.h
@@ -32,7 +32,7 @@
 /*
  * Each page of XLOG file has a header like this:
  */
-#define XLOG_PAGE_MAGIC 0xD11E	/* can be used as WAL version indicator */
+#define XLOG_PAGE_MAGIC 0xD11F	/* can be used as WAL version indicator */
 
 typedef struct XLogPageHeaderData
 {
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 2e10e3c814d..93b7816c09c 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -310,6 +310,15 @@ extern void PreventCommandIfReadOnly(const char *cmdname);
 extern void PreventCommandIfParallelMode(const char *cmdname);
 extern void PreventCommandDuringRecovery(const char *cmdname);
 
+/* in replication/snapbuild.c */
+
+/*
+ * Keep track of whether logical decoding in this backend promised not to
+ * access shared catalogs, as a safety check.  This is checked by genam.c when
+ * a catalog scan takes place to verify that no shared catalogs are accessed.
+ */
+extern bool accessSharedCatalogsInDecoding;
+
 /*****************************************************************************
  *	  pdir.h --																 *
  *			POSTGRES directory path definitions.                             *
diff --git a/src/include/replication/output_plugin.h b/src/include/replication/output_plugin.h
index 842fcde67f9..917f3cff232 100644
--- a/src/include/replication/output_plugin.h
+++ b/src/include/replication/output_plugin.h
@@ -27,6 +27,7 @@ typedef struct OutputPluginOptions
 {
 	OutputPluginOutputType output_type;
 	bool		receive_rewrites;
+	bool		need_shared_catalogs;
 } OutputPluginOptions;
 
 /*
diff --git a/src/include/replication/snapbuild.h b/src/include/replication/snapbuild.h
index a22a83a2f23..d02530a912a 100644
--- a/src/include/replication/snapbuild.h
+++ b/src/include/replication/snapbuild.h
@@ -92,7 +92,8 @@ extern void SnapBuildProcessNewCid(SnapBuild *builder, TransactionId xid,
 								   XLogRecPtr lsn,
 								   xl_heap_new_cid *xlrec);
 extern void SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn,
-										 xl_running_xacts *running);
+										 xl_running_xacts *running,
+										 bool db_specific);
 extern void SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn);
 
 extern bool SnapBuildSnapshotExists(XLogRecPtr lsn);
diff --git a/src/include/storage/procarray.h b/src/include/storage/procarray.h
index d718a5b542f..ec89c448220 100644
--- a/src/include/storage/procarray.h
+++ b/src/include/storage/procarray.h
@@ -47,7 +47,7 @@ extern bool ProcArrayInstallImportedXmin(TransactionId xmin,
 										 VirtualTransactionId *sourcevxid);
 extern bool ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc);
 
-extern RunningTransactions GetRunningTransactionData(void);
+extern RunningTransactions GetRunningTransactionData(Oid dbid);
 
 extern bool TransactionIdIsInProgress(TransactionId xid);
 extern TransactionId GetOldestNonRemovableTransactionId(Relation rel);
diff --git a/src/include/storage/standby.h b/src/include/storage/standby.h
index 6a314c693cd..8715c08e94f 100644
--- a/src/include/storage/standby.h
+++ b/src/include/storage/standby.h
@@ -126,6 +126,7 @@ typedef enum
 
 typedef struct RunningTransactionsData
 {
+	Oid			dbid;			/* only track xacts in this database */
 	int			xcnt;			/* # of xact ids in xids[] */
 	int			subxcnt;		/* # of subxact ids in xids[] */
 	subxids_array_status subxid_status;
@@ -143,7 +144,7 @@ typedef RunningTransactionsData *RunningTransactions;
 extern void LogAccessExclusiveLock(Oid dbOid, Oid relOid);
 extern void LogAccessExclusiveLockPrepare(void);
 
-extern XLogRecPtr LogStandbySnapshot(void);
+extern XLogRecPtr LogStandbySnapshot(Oid dbid);
 extern void LogStandbyInvalidations(int nmsgs, SharedInvalidationMessage *msgs,
 									bool relcacheInitFileInval);
 
diff --git a/src/include/storage/standbydefs.h b/src/include/storage/standbydefs.h
index 231d251fd51..e75b7078766 100644
--- a/src/include/storage/standbydefs.h
+++ b/src/include/storage/standbydefs.h
@@ -46,6 +46,7 @@ typedef struct xl_standby_locks
  */
 typedef struct xl_running_xacts
 {
+	Oid			dbid;			/* only track xacts in this database */
 	int			xcnt;			/* # of xact ids in xids[] */
 	int			subxcnt;		/* # of subxact ids in xids[] */
 	bool		subxid_overflow;	/* snapshot overflowed, subxids missing */
-- 
2.47.3


--qs77q6fpwolne4ds
Content-Type: text/x-diff; charset=utf-8
Content-Disposition: attachment;
	filename="v56-0002-Reserve-replication-slots-specifically-for-REPAC.patch"



^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread

* [PATCH] Lock upgrade without deadlocks.
@ 2026-04-17 13:13  Antonin Houska <[email protected]>
  0 siblings, 0 replies; 200+ messages in thread

From: Antonin Houska @ 2026-04-17 13:13 UTC (permalink / raw)

For REPACK (CONCURRENTLY), it's essential to upgrade ShareUpdateExclusive lock
to AccessExclusiveLock for the final stage of table processing. However, we
cannot release the earlier before requesting the latter, because if another
process ran DDL on the table in between, REPACK would have to cancel its
transaction, and waste possibly a lot of work. Without releasing
ShareUpdateExclusive lock, the upgrade can cause deadlock, and it's not really
deterministic whether REPACK or the other process will get canceled.

This patch adds a new field to the LOCK structure which tells which process is
trying to upgrade the lock - this field is set right before the upgrade. All
processes waiting in the lock queue are then woken up to check the flag.

When another process tries to get the lock after that, it checks this field,
and if it's set, it checks for deadlocks even if deadlock timeout hasn't
expired yet. If the process is already in the lock's queue and sleeping, the
lock upgrading process wakes it up so it checks the flag immediately.

At the time the upgrading process starts to sleep, all the other process
should be aware that they need to check for deadlock during their lock
acquisitions, so the upgrading process does not have to check for deadlocks
when it gets woken up. Thus it should never receive deadlock error report.
---
 src/backend/commands/repack.c   |   4 +-
 src/backend/storage/lmgr/lmgr.c |  35 ++++++++---
 src/backend/storage/lmgr/lock.c | 107 +++++++++++++++++++++++++++++++-
 src/backend/storage/lmgr/proc.c |  32 +++++++---
 src/include/storage/lmgr.h      |   1 +
 src/include/storage/lock.h      |   5 +-
 src/include/storage/proc.h      |   4 +-
 7 files changed, 166 insertions(+), 22 deletions(-)

diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 58e3867246f..310b2a65099 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3055,8 +3055,10 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap,
 	/*
 	 * Acquire AccessExclusiveLock on the table, its TOAST relation (if there
 	 * is one), all its indexes, so that we can swap the files.
+	 *
+	 * TODO The same for indexes and TOAST?
 	 */
-	LockRelationOid(old_table_oid, AccessExclusiveLock);
+	LockRelationOidUpgrade(old_table_oid, AccessExclusiveLock);
 
 	/*
 	 * Lock all indexes now, not only the clustering one: all indexes need to
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index 2ccf7237fee..0b103485d54 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -58,6 +58,8 @@ typedef struct XactLockTableWaitInfo
 	const ItemPointerData *ctid;
 } XactLockTableWaitInfo;
 
+static void LockRelationOidCommon(Oid relid, LOCKMODE lockmode,
+								  bool isUpgrade);
 static void XactLockTableWaitErrorCb(void *arg);
 
 /*
@@ -105,6 +107,21 @@ SetLocktagRelationOid(LOCKTAG *tag, Oid relid)
  */
 void
 LockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, false);
+}
+
+/*
+ * Like above, but upgrade an existing lock.
+ */
+void
+LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode)
+{
+	LockRelationOidCommon(relid, lockmode, true);
+}
+
+static void
+LockRelationOidCommon(Oid relid, LOCKMODE lockmode, bool isUpgrade)
 {
 	LOCKTAG		tag;
 	LOCALLOCK  *locallock;
@@ -113,7 +130,7 @@ LockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, isUpgrade);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages, so that we
@@ -157,7 +174,7 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 	SetLocktagRelationOid(&tag, relid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -191,7 +208,7 @@ LockRelationId(LockRelId *relid, LOCKMODE lockmode)
 	SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -254,7 +271,7 @@ LockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock,
-							  false);
+							  false, false);
 
 	/*
 	 * Now that we have the lock, check for invalidation messages; see notes
@@ -286,7 +303,7 @@ ConditionalLockRelation(Relation relation, LOCKMODE lockmode)
 						 relation->rd_lockInfo.lockRelId.relId);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -591,7 +608,7 @@ ConditionalLockTuple(Relation relation, const ItemPointerData *tid, LOCKMODE loc
 					  ItemPointerGetOffsetNumber(tid));
 
 	return (LockAcquireExtended(&tag, lockmode, false, true, true, NULL,
-								logLockFailure) != LOCKACQUIRE_NOT_AVAIL);
+								logLockFailure, false) != LOCKACQUIRE_NOT_AVAIL);
 }
 
 /*
@@ -749,7 +766,7 @@ ConditionalXactLockTableWait(TransactionId xid, bool logLockFailure)
 		SET_LOCKTAG_TRANSACTION(tag, xid);
 
 		if (LockAcquireExtended(&tag, ShareLock, false, true, true, NULL,
-								logLockFailure)
+								logLockFailure, false)
 			== LOCKACQUIRE_NOT_AVAIL)
 			return false;
 
@@ -1043,7 +1060,7 @@ ConditionalLockDatabaseObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
@@ -1123,7 +1140,7 @@ ConditionalLockSharedObject(Oid classid, Oid objid, uint16 objsubid,
 					   objsubid);
 
 	res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock,
-							  false);
+							  false, false);
 
 	if (res == LOCKACQUIRE_NOT_AVAIL)
 		return false;
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index c221fe96889..bcbbf36c091 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -810,7 +810,7 @@ LockAcquire(const LOCKTAG *locktag,
 			bool dontWait)
 {
 	return LockAcquireExtended(locktag, lockmode, sessionLock, dontWait,
-							   true, NULL, false);
+							   true, NULL, false, false);
 }
 
 /*
@@ -829,6 +829,9 @@ LockAcquire(const LOCKTAG *locktag,
  *
  * logLockFailure indicates whether to log details when a lock acquisition
  * fails with dontWait = true.
+ *
+ * isUpgrade should be true if the backend already holds this lock in a lower
+ * mode.
  */
 LockAcquireResult
 LockAcquireExtended(const LOCKTAG *locktag,
@@ -837,7 +840,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 					bool dontWait,
 					bool reportMemoryError,
 					LOCALLOCK **locallockp,
-					bool logLockFailure)
+					bool logLockFailure,
+					bool isUpgrade)
 {
 	LOCKMETHODID lockmethodid = locktag->locktag_lockmethodid;
 	LockMethod	lockMethodTable;
@@ -1119,7 +1123,8 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		 * case, because JoinWaitQueue() may discover that we can acquire the
 		 * lock immediately after all.
 		 */
-		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait);
+		waitResult = JoinWaitQueue(locallock, lockMethodTable, dontWait,
+								   isUpgrade);
 	}
 
 	if (waitResult == PROC_WAIT_STATUS_ERROR)
@@ -1225,8 +1230,85 @@ LockAcquireExtended(const LOCKTAG *locktag,
 		Assert(!dontWait);
 		PROCLOCK_PRINT("LockAcquire: sleeping on lock", proclock);
 		LOCK_PRINT("LockAcquire: sleeping on lock", lock, lockmode);
+
+		/*
+		 * Lock upgrade can introduce deadlock, therefore enforce special
+		 * behavior of other processes that deal with this lock. In
+		 * particular, any waiter that sees upgradedBy set is expected to
+		 * perform deadlock check as soon as it's woken up. (Since we are
+		 * already in the queue, the deadlock detector has all the information
+		 * it needs.)
+		 */
+		if (isUpgrade)
+		{
+			dlist_iter	iter;
+
+			/*
+			 * There should not be multiple upgrades at the same time. XXX
+			 * ERROR ?
+			 */
+			Assert(lock->upgradedBy == NULL);
+
+			lock->upgradedBy = MyProc;
+
+			/*
+			 * Now, before I sleep myself, wake up all the existing waiters
+			 * (except for me), so they check for deadlock.
+			 */
+			dclist_foreach(iter, &lock->waitProcs)
+			{
+				PGPROC	   *proc = dlist_container(PGPROC, waitLink,
+												   iter.cur);
+
+				if (proc != MyProc)
+					SetLatch(&proc->procLatch);
+			}
+		}
 		LWLockRelease(partitionLock);
 
+		if (!isUpgrade)
+		{
+			HASH_SEQ_STATUS status;
+			LOCALLOCK	*llock;
+			bool	check_deadlock = false;
+
+			/*
+			 * Likewise, new waiters should check the isUpgrade flag and
+			 * engage deadlock detector if needed. The lock being acquired
+			 * should already be in the hash.
+			 *
+			 * Note that, to make deadlock detection happen as soon as
+			 * possible (i.e. regardless deadlock timeout), we check the other
+			 * locks too. XXX There might be ways to get into the deadlock
+			 * indirectly, but that needs more analysis. As long as the
+			 * upgrading process skips deadlock detection altogether, the
+			 * worst consequence of missing some lock wait cycle here is that
+			 * the other process will be kicked-off no sooner than after its
+			 * deadlock timeout has elapsed. (Which in turn means that the
+			 * lock upgrade will take longer than expected.)
+			 */
+			hash_seq_init(&status, LockMethodLocalHash);
+			while ((llock = (LOCALLOCK *) hash_seq_search(&status)) != NULL)
+			{
+				LOCK	*mylock = llock->lock;
+
+				if (mylock && mylock->upgradedBy)
+				{
+					check_deadlock = true;
+					break;
+				}
+			}
+			/* Check for deadlock if needed. */
+			if (check_deadlock)
+			{
+				DeadLockState	deadlock_state;
+
+				deadlock_state = CheckDeadLock();
+				if (deadlock_state == DS_HARD_DEADLOCK)
+					DeadLockReport();
+			}
+		}
+
 		waitResult = WaitOnLock(locallock, owner);
 
 		/*
@@ -1245,6 +1327,18 @@ LockAcquireExtended(const LOCKTAG *locktag,
 			DeadLockReport();
 			/* DeadLockReport() will not return */
 		}
+
+		/*
+		 * If finishing the lock upgrade, we should not get into a deadlock
+		 * anymore, so let others know that they do not have to care either.
+		 */
+		if (isUpgrade)
+		{
+			LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+			Assert(lock->upgradedBy != NULL);
+			lock->upgradedBy = NULL;
+			LWLockRelease(partitionLock);
+		}
 	}
 	else
 		LWLockRelease(partitionLock);
@@ -1322,6 +1416,7 @@ SetupLockInTable(LockMethod lockMethodTable, PGPROC *proc,
 		lock->nGranted = 0;
 		MemSet(lock->requested, 0, sizeof(int) * MAX_LOCKMODES);
 		MemSet(lock->granted, 0, sizeof(int) * MAX_LOCKMODES);
+		lock->upgradedBy = NULL;
 		LOCK_PRINT("LockAcquire: new", lock, lockmode);
 	}
 	else
@@ -1768,6 +1863,12 @@ CleanUpLock(LOCK *lock, PROCLOCK *proclock,
 			elog(PANIC, "proclock table corrupted");
 	}
 
+	/*
+	 * Was this backend upgrading the lock?
+	 */
+	if (lock->upgradedBy == MyProc)
+		lock->upgradedBy = NULL;
+
 	if (lock->nRequested == 0)
 	{
 		/*
diff --git a/src/backend/storage/lmgr/proc.c b/src/backend/storage/lmgr/proc.c
index 1ac25068d62..7b31370db70 100644
--- a/src/backend/storage/lmgr/proc.c
+++ b/src/backend/storage/lmgr/proc.c
@@ -95,7 +95,6 @@ static volatile sig_atomic_t got_deadlock_timeout;
 static void RemoveProcFromArray(int code, Datum arg);
 static void ProcKill(int code, Datum arg);
 static void AuxiliaryProcKill(int code, Datum arg);
-static DeadLockState CheckDeadLock(void);
 
 
 /*
@@ -1143,7 +1142,8 @@ AuxiliaryPidGetProc(int pid)
  * NOTES: The process queue is now a priority queue for locking.
  */
 ProcWaitStatus
-JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
+JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait,
+			  bool isUpgrade)
 {
 	LOCKMODE	lockmode = locallock->tag.mode;
 	LOCK	   *lock = locallock->lock;
@@ -1230,8 +1230,12 @@ JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
 			/* Must he wait for me? */
 			if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
 			{
-				/* Must I wait for him ? */
-				if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
+				/*
+				 * Must I wait for him? I don't want a deadlock during lock
+				 * upgrade - other processes should fail on it.
+				 */
+				if ((lockMethodTable->conflictTab[lockmode] & proc->heldLocks) &&
+					!isUpgrade)
 				{
 					/*
 					 * Yes, so we have a deadlock.  Easiest way to clean up
@@ -1461,8 +1465,22 @@ ProcSleep(LOCALLOCK *locallock)
 			(void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
 							 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
 			ResetLatch(MyLatch);
-			/* check for deadlocks first, as that's probably log-worthy */
-			if (got_deadlock_timeout)
+			/*
+			 * Check for deadlocks first, as that's probably log-worthy.
+			 *
+			 * Do not wait for the timeout if the lock is being upgraded since
+			 * the risk of deadlock is higher now. However, do not check for
+			 * deadlock if the lock is being upgraded by this process - other
+			 * processes should take care.
+			 *
+			 * TODO Possible optimization: if this is the only lock of the
+			 * backend and if it did not have any weaker lock on the table so
+			 * far, it should be safe to skip the deadlock check. However, to
+			 * evaluate the situation, we need to take fast-path locks into
+			 * account.
+			 */
+			if ((got_deadlock_timeout || lock->upgradedBy) &&
+				lock->upgradedBy != MyProc)
 			{
 				deadlock_state = CheckDeadLock();
 				got_deadlock_timeout = false;
@@ -1819,7 +1837,7 @@ ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
  * not, just return.  If we have a real deadlock, remove ourselves from the
  * lock's wait queue.
  */
-static DeadLockState
+DeadLockState
 CheckDeadLock(void)
 {
 	int			i;
diff --git a/src/include/storage/lmgr.h b/src/include/storage/lmgr.h
index 2a985ce5e15..eddbe1e8622 100644
--- a/src/include/storage/lmgr.h
+++ b/src/include/storage/lmgr.h
@@ -38,6 +38,7 @@ extern void RelationInitLockInfo(Relation relation);
 
 /* Lock a relation */
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
+extern void LockRelationOidUpgrade(Oid relid, LOCKMODE lockmode);
 extern void LockRelationId(LockRelId *relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index ee3cb1dc203..080b69eddfd 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -131,6 +131,7 @@ typedef const LockMethodData *LockMethod;
  * nRequested -- total requested locks of all types.
  * granted -- count of each lock type currently granted on the lock.
  * nGranted -- total granted locks of all types.
+ * upgrader -- process that intends to upgrade the lock
  *
  * Note: these counts count 1 for each backend.  Internally to a backend,
  * there may be multiple grabs on a particular lock, but this is not reflected
@@ -150,6 +151,7 @@ typedef struct LOCK
 	int			nRequested;		/* total of requested[] array */
 	int			granted[MAX_LOCKMODES]; /* counts of granted locks */
 	int			nGranted;		/* total of granted[] array */
+	PGPROC		*upgradedBy;	/* is lock being upgraded by this process? */
 } LOCK;
 
 #define LOCK_LOCKMETHOD(lock) ((LOCKMETHODID) (lock).tag.locktag_lockmethodid)
@@ -390,7 +392,8 @@ extern LockAcquireResult LockAcquireExtended(const LOCKTAG *locktag,
 											 bool dontWait,
 											 bool reportMemoryError,
 											 LOCALLOCK **locallockp,
-											 bool logLockFailure);
+											 bool logLockFailure,
+											 bool isUpgrade);
 extern void AbortStrongLockAcquire(void);
 extern void MarkLockClear(LOCALLOCK *locallock);
 extern bool LockRelease(const LOCKTAG *locktag,
diff --git a/src/include/storage/proc.h b/src/include/storage/proc.h
index 3e1d1fad5f9..c4bc3a24044 100644
--- a/src/include/storage/proc.h
+++ b/src/include/storage/proc.h
@@ -563,10 +563,12 @@ extern bool HaveNFreeProcs(int n, int *nfree);
 extern void ProcReleaseLocks(bool isCommit);
 
 extern ProcWaitStatus JoinWaitQueue(LOCALLOCK *locallock,
-									LockMethod lockMethodTable, bool dontWait);
+									LockMethod lockMethodTable, bool dontWait,
+									bool isUpgrade);
 extern ProcWaitStatus ProcSleep(LOCALLOCK *locallock);
 extern void ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus);
 extern void ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock);
+extern DeadLockState CheckDeadLock(void);
 extern void CheckDeadLockAlert(void);
 extern void LockErrorCleanup(void);
 extern void GetLockHoldersAndWaiters(LOCALLOCK *locallock,
-- 
2.47.3


--=-=-=--





^ permalink  raw  reply  [nested|flat] 200+ messages in thread