Hi, On Wed, Sep 16, 2026 at 2:02 PM Masahiko Sawada <[email protected]> wrote: > > > Sawada-san, does that mean that on the back branches, even for the > > case where the concerned subtransaction is committing while the slot > > is still held (a scenario we don't know can happen), we would release > > the slot and clean up MyReplicationSlotSubId? Is my understanding > > correct? > > I don't think we should release the slot at subxact commit. I think > it's better to leave it to the caller as it might release the slot > afterward. Another problem is that nothing tests this case. > > Please refer to the attached patch that can be applied on v16 patch > and implements my idea. It adds additional regression tests too.
Thanks for sharing the diff. I apologize for going back and forth on the subtransaction commit hand-off. After looking at it, here is what I have. The slot acquired in a subtransaction gets cleaned up only if that subtransaction is aborted, which is the reported bug and easily reachable from SQL today. MyReplicationSlotSubId is not a stale value within the transaction but the owner's subtransaction id. However, across transactions it becomes stale and can get reused, so it is reset at the subtransaction commit path, preventing MyReplicationSlotSubId from going stale for users carrying the slot across transactions (if any). In core code, the owning subtransaction always releases the slot before it ends, and the release clears the id. No hand-off, no warning on subtransaction commit. Please find the attached patch for HEAD. I dropped the two unnecessary header file inclusions added in the diff but otherwise took it as-is. If it looks good, I can prepare the patches for all supported branches. -- Bharath Rupireddy Amazon Web Services: https://aws.amazon.com
From 59fb23f58b3732ed49344889dee7a24c68bd52df Mon Sep 17 00:00:00 2001 From: Bharath Rupireddy <[email protected]> Date: Thu, 17 Sep 2026 01:12:24 +0000 Subject: [PATCH v17] Fix replication slot leak on error caught in a subtransaction. The SQL-callable replication slot functions, such as pg_replication_slot_advance(), acquire a slot and release it before returning. When one of them throws an error, the slot is normally released by the top-level error handler (the sigsetjmp block in PostgresMain()). However, PL/pgSQL, PL/Perl, PL/Python and PL/Tcl run their error-handling block in an internal subtransaction. When an error is raised there and a handler matches, they abort the subtransaction and catch the error without re-throwing it, so the top-level handler is never reached. Nothing on the (sub)transaction abort path releases the slot either, so it is left acquired after the caught error. As a result, the next slot operation in the session fails an assertion, or in a non-assertion build silently overwrites the reference and leaks the slot, which keeps holding back WAL removal and the catalog xmin (blocking vacuum) and can no longer be acquired. An error that does reach the top-level handler, such as a plain ROLLBACK TO SAVEPOINT or a PL/pgSQL exception handler whose conditions do not match and which then re-throws, is unaffected. Fix this by recording the subtransaction that acquires the slot and releasing the slot when that subtransaction aborts, the same way other subtransaction-scoped resources are released, for example in AtEOSubXact_LargeObject() and AtEOSubXact_Files(). The slot cannot simply be released on every subtransaction abort, because logical decoding starts and aborts an internal transaction or subtransaction for each decoded transaction while holding the slot; those internal aborts are nested below the acquiring subtransaction and so are left alone. The recorded id is the id of the subtransaction that owns the slot, and it is cleared when the slot is released. It is not handed to the parent when that subtransaction commits with the slot still held. No slot function in the core code returns with the slot held, and a slot left unreleased by mistake cannot be told apart from one the caller means to release later, so the id is forgotten there and the slot is left to the caller. Note that the subtransaction abort path only releases the slot; it does not drop the session's temporary slots the way the top-level error handler does. An error caught within a subtransaction is normally meant to be handled so the session continues, unlike a top-level error, so temporary slots are left in place. This matches the behavior that predates this fix, where a temporary slot is dropped only at session end or on a top-level error, and is simpler to reason about. The documentation is adjusted accordingly. Backpatch to all supported branches, as the leak is reachable from SQL. Reported-by: Satya Narlapuram <[email protected]> Author: Bharath Rupireddy <[email protected]> Co-authored-by: Satya Narlapuram <[email protected]> Co-authored-by: Masahiko Sawada <[email protected]> Reviewed-by: Fujii Masao <[email protected]> Reviewed-by: shveta malik <[email protected]> Reviewed-by: Hou Zhijie <[email protected]> Reviewed-by: Kyotaro Horiguchi <[email protected]> Reviewed-by: Ashutosh Sharma <[email protected]> Reviewed-by: Amit Kapila <[email protected]> Discussion: https://postgr.es/m/CAHg+QDeuf9tCq3ce=kgFMJP0m=PZC+wi6B=yS+7V0vNXjLS31w@mail.gmail.com Backpatch-through: 14 --- contrib/test_decoding/expected/slot.out | 89 +++++++++++++++++++++++++ contrib/test_decoding/sql/slot.sql | 45 +++++++++++++ doc/src/sgml/func/func-admin.sgml | 10 ++- doc/src/sgml/protocol.sgml | 5 +- doc/src/sgml/system-views.sgml | 5 +- src/backend/access/transam/xact.c | 2 + src/backend/replication/slot.c | 66 ++++++++++++++++++ src/backend/tcop/postgres.c | 4 +- src/include/replication/slot.h | 1 + 9 files changed, 222 insertions(+), 5 deletions(-) diff --git a/contrib/test_decoding/expected/slot.out b/contrib/test_decoding/expected/slot.out index cf0136455b0..f7906f1e4c4 100644 --- a/contrib/test_decoding/expected/slot.out +++ b/contrib/test_decoding/expected/slot.out @@ -469,3 +469,92 @@ SELECT pg_drop_replication_slot('physical_slot'); (1 row) +-- A slot function that errors out must still release the slot, otherwise the +-- next slot operation in the session fails an assertion or leaks the slot. +-- Advancing a freshly created slot to a low LSN always errors. +SELECT 'init' FROM pg_create_logical_replication_slot('regress_subxact_slot', 'test_decoding'); + ?column? +---------- + init +(1 row) + +-- Error raised inside a PL/pgSQL block with an EXCEPTION clause is caught in a +-- subtransaction; the slot must still be released. +DO $$ +BEGIN + PERFORM pg_replication_slot_advance('regress_subxact_slot', '0/1'); +EXCEPTION WHEN OTHERS THEN + RAISE NOTICE 'caught SQLSTATE %', SQLSTATE; +END $$; +NOTICE: caught SQLSTATE 55000 +SELECT active FROM pg_replication_slots WHERE slot_name = 'regress_subxact_slot'; + active +-------- + f +(1 row) + +SELECT count(*) >= 0 AS peek_ok + FROM pg_logical_slot_peek_changes('regress_subxact_slot', NULL, NULL); + peek_ok +--------- + t +(1 row) + +-- The EXCEPTION clause does not match, so the error is not caught, but the +-- slot is still released. Show only the SQLSTATE for stable output. +\set VERBOSITY sqlstate +DO $$ +BEGIN + PERFORM pg_replication_slot_advance('regress_subxact_slot', '0/1'); +EXCEPTION WHEN division_by_zero THEN + RAISE NOTICE 'unreachable'; +END $$; +ERROR: 55000 +\set VERBOSITY default +SELECT active FROM pg_replication_slots WHERE slot_name = 'regress_subxact_slot'; + active +-------- + f +(1 row) + +SELECT count(*) >= 0 AS peek_ok + FROM pg_logical_slot_peek_changes('regress_subxact_slot', NULL, NULL); + peek_ok +--------- + t +(1 row) + +SELECT pg_drop_replication_slot('regress_subxact_slot'); + pg_drop_replication_slot +-------------------------- + +(1 row) + +-- Unlike a top-level error, an error caught in a subtransaction releases the +-- slot but leaves the session's temporary slots in place. +SELECT 'init' FROM pg_create_logical_replication_slot('regress_subxact_temp_slot', 'test_decoding', true); + ?column? +---------- + init +(1 row) + +DO $$ +BEGIN + PERFORM pg_replication_slot_advance('regress_subxact_temp_slot', '0/1'); +EXCEPTION WHEN OTHERS THEN + RAISE NOTICE 'caught SQLSTATE %', SQLSTATE; +END $$; +NOTICE: caught SQLSTATE 55000 +SELECT count(*) = 1 AS temp_slot_kept + FROM pg_replication_slots WHERE slot_name = 'regress_subxact_temp_slot'; + temp_slot_kept +---------------- + t +(1 row) + +SELECT pg_drop_replication_slot('regress_subxact_temp_slot'); + pg_drop_replication_slot +-------------------------- + +(1 row) + diff --git a/contrib/test_decoding/sql/slot.sql b/contrib/test_decoding/sql/slot.sql index 50a950341fc..bd292c41e9b 100644 --- a/contrib/test_decoding/sql/slot.sql +++ b/contrib/test_decoding/sql/slot.sql @@ -193,3 +193,48 @@ SELECT pg_drop_replication_slot('failover_true_slot'); SELECT pg_drop_replication_slot('failover_false_slot'); SELECT pg_drop_replication_slot('failover_default_slot'); SELECT pg_drop_replication_slot('physical_slot'); + +-- A slot function that errors out must still release the slot, otherwise the +-- next slot operation in the session fails an assertion or leaks the slot. +-- Advancing a freshly created slot to a low LSN always errors. +SELECT 'init' FROM pg_create_logical_replication_slot('regress_subxact_slot', 'test_decoding'); + +-- Error raised inside a PL/pgSQL block with an EXCEPTION clause is caught in a +-- subtransaction; the slot must still be released. +DO $$ +BEGIN + PERFORM pg_replication_slot_advance('regress_subxact_slot', '0/1'); +EXCEPTION WHEN OTHERS THEN + RAISE NOTICE 'caught SQLSTATE %', SQLSTATE; +END $$; +SELECT active FROM pg_replication_slots WHERE slot_name = 'regress_subxact_slot'; +SELECT count(*) >= 0 AS peek_ok + FROM pg_logical_slot_peek_changes('regress_subxact_slot', NULL, NULL); + +-- The EXCEPTION clause does not match, so the error is not caught, but the +-- slot is still released. Show only the SQLSTATE for stable output. +\set VERBOSITY sqlstate +DO $$ +BEGIN + PERFORM pg_replication_slot_advance('regress_subxact_slot', '0/1'); +EXCEPTION WHEN division_by_zero THEN + RAISE NOTICE 'unreachable'; +END $$; +\set VERBOSITY default +SELECT active FROM pg_replication_slots WHERE slot_name = 'regress_subxact_slot'; +SELECT count(*) >= 0 AS peek_ok + FROM pg_logical_slot_peek_changes('regress_subxact_slot', NULL, NULL); +SELECT pg_drop_replication_slot('regress_subxact_slot'); + +-- Unlike a top-level error, an error caught in a subtransaction releases the +-- slot but leaves the session's temporary slots in place. +SELECT 'init' FROM pg_create_logical_replication_slot('regress_subxact_temp_slot', 'test_decoding', true); +DO $$ +BEGIN + PERFORM pg_replication_slot_advance('regress_subxact_temp_slot', '0/1'); +EXCEPTION WHEN OTHERS THEN + RAISE NOTICE 'caught SQLSTATE %', SQLSTATE; +END $$; +SELECT count(*) = 1 AS temp_slot_kept + FROM pg_replication_slots WHERE slot_name = 'regress_subxact_temp_slot'; +SELECT pg_drop_replication_slot('regress_subxact_temp_slot'); diff --git a/doc/src/sgml/func/func-admin.sgml b/doc/src/sgml/func/func-admin.sgml index 64b0e7bb972..5b250f46ed2 100644 --- a/doc/src/sgml/func/func-admin.sgml +++ b/doc/src/sgml/func/func-admin.sgml @@ -1051,7 +1051,10 @@ postgres=# SELECT '0/0'::pg_lsn + pd.segment_number * ps.setting::int + :offset parameter, <parameter>temporary</parameter>, when set to true, specifies that the slot should not be permanently stored to disk and is only meant for use by the current session. Temporary slots are also - released upon any error. This function corresponds + dropped on any error. An error raised and caught in a + subtransaction, for example by a + <application>PL/pgSQL</application> exception block, does not + drop them. This function corresponds to the replication protocol command <literal>CREATE_REPLICATION_SLOT ... PHYSICAL</literal>. </para></entry> @@ -1091,7 +1094,10 @@ postgres=# SELECT '0/0'::pg_lsn + pd.segment_number * ps.setting::int + :offset parameter, <parameter>temporary</parameter>, when set to true, specifies that the slot should not be permanently stored to disk and is only meant for use by the current session. Temporary slots are also - released upon any error. The optional fourth parameter, + dropped on any error. An error raised and caught in a + subtransaction, for example by a + <application>PL/pgSQL</application> exception block, does not + drop them. The optional fourth parameter, <parameter>twophase</parameter>, when set to true, specifies that the decoding of prepared transactions is enabled for this slot. The optional fifth parameter, diff --git a/doc/src/sgml/protocol.sgml b/doc/src/sgml/protocol.sgml index 8ced69f283b..ae1a04b3454 100644 --- a/doc/src/sgml/protocol.sgml +++ b/doc/src/sgml/protocol.sgml @@ -2371,7 +2371,10 @@ psql "dbname=postgres replication=database" -c "IDENTIFY_SYSTEM;" <para> Specify that this replication slot is a temporary one. Temporary slots are not saved to disk and are automatically dropped on error - or when the session has finished. + or when the session has finished. An error raised and caught in a + subtransaction, for example by a + <application>PL/pgSQL</application> exception block, does not drop + them. </para> </listitem> </varlistentry> diff --git a/doc/src/sgml/system-views.sgml b/doc/src/sgml/system-views.sgml index 77202e2c765..efefc43e6fc 100644 --- a/doc/src/sgml/system-views.sgml +++ b/doc/src/sgml/system-views.sgml @@ -2894,7 +2894,10 @@ SELECT * FROM pg_locks pl LEFT JOIN pg_prepared_xacts ppx <para> True if this is a temporary replication slot. Temporary slots are not saved to disk and are automatically dropped on error or when - the session has finished. + the session has finished. An error raised and caught in a + subtransaction, for example by a + <application>PL/pgSQL</application> exception block, does not drop + them. </para></entry> </row> diff --git a/src/backend/access/transam/xact.c b/src/backend/access/transam/xact.c index aca92507ebd..3efb09bbb02 100644 --- a/src/backend/access/transam/xact.c +++ b/src/backend/access/transam/xact.c @@ -5203,6 +5203,7 @@ CommitSubTransaction(void) s->parent->curTransactionOwner); AtEOSubXact_LargeObject(true, s->subTransactionId, s->parent->subTransactionId); + AtEOSubXact_ReplicationSlot(true, s->subTransactionId); AtSubCommit_Notify(); CallSubXactCallbacks(SUBXACT_EVENT_COMMIT_SUB, s->subTransactionId, @@ -5380,6 +5381,7 @@ AbortSubTransaction(void) s->parent->curTransactionOwner); AtEOSubXact_LargeObject(false, s->subTransactionId, s->parent->subTransactionId); + AtEOSubXact_ReplicationSlot(false, s->subTransactionId); AtSubAbort_Notify(); /* Advertise the fact that we aborted in pg_xact. */ diff --git a/src/backend/replication/slot.c b/src/backend/replication/slot.c index 63ce6d27885..45766541bed 100644 --- a/src/backend/replication/slot.c +++ b/src/backend/replication/slot.c @@ -59,6 +59,7 @@ #include "utils/builtins.h" #include "utils/guc_hooks.h" #include "utils/injection_point.h" +#include "utils/snapmgr.h" #include "utils/varlena.h" #include "utils/wait_event.h" @@ -157,6 +158,13 @@ const ShmemCallbacks ReplicationSlotsShmemCallbacks = { /* My backend's replication slot in the shared memory array */ ReplicationSlot *MyReplicationSlot = NULL; +/* + * Subxact that acquired MyReplicationSlot, or invalid if none is held + * or it was acquired with no transaction in progress (as a walsender does). + * Used to release the slot when that subxact aborts. + */ +static SubTransactionId MyReplicationSlotSubId = InvalidSubTransactionId; + /* GUC variables */ int max_replication_slots = 10; /* the maximum number of replication * slots */ @@ -519,6 +527,7 @@ ReplicationSlotCreate(const char *name, bool db_specific, slot->active_proc = MyProcNumber; SpinLockRelease(&slot->mutex); MyReplicationSlot = slot; + MyReplicationSlotSubId = GetCurrentSubTransactionId(); LWLockRelease(ReplicationSlotControlLock); @@ -724,6 +733,7 @@ retry: /* We made this slot active, so it's ours now. */ MyReplicationSlot = s; + MyReplicationSlotSubId = GetCurrentSubTransactionId(); /* * We need to check for invalidation after making the slot ours to avoid @@ -848,6 +858,61 @@ ReplicationSlotRelease(void) pfree(slotname); } + + /* The slot is no longer acquired in any subxact. */ + MyReplicationSlotSubId = InvalidSubTransactionId; +} + +/* + * At subxact end, release the replication slot if the subtransaction + * where the slot was acquired is aborted. + */ +void +AtEOSubXact_ReplicationSlot(bool isCommit, SubTransactionId mySubid) +{ + /* Nothing to do unless the slot was acquired in this subxact. */ + if (MyReplicationSlotSubId != mySubid) + return; + + /* + * The subxact that acquired the slot is committing with the slot still + * held. No slot function in the core code does that today, though an + * extension might. A slot left unreleased by mistake cannot be told apart + * from one the caller means to release later. Leave it to the caller. + */ + if (isCommit) + { + MyReplicationSlotSubId = InvalidSubTransactionId; + return; + } + + /* + * We must not get here while decoding is running. Decoding starts and + * aborts an internal (sub)transaction while holding the slot, for each + * decoded transaction (ReorderBufferProcessTXN()) and when executing + * invalidations (ReorderBufferImmediateInvalidation()). However, those + * subtransactions are always nested below the one that acquired the slot, + * so their subtransaction ids are deeper and do not match here. Decoding + * also runs with a historic snapshot set up, so assert that it is not. + */ + Assert(!HistoricSnapshotActive()); + + /* + * The aborting subxact is the one that acquired the slot, so the slot is + * still held and must be released. MyReplicationSlotSubId is set only + * when a slot is held and cleared when it is released, so a matching + * subxact id means the slot is ours. + * + * We only release the slot here and do not drop the session's temporary + * slots, unlike the top-level error handler in PostgresMain(). An error + * caught within a subtransaction, for example by a PL/pgSQL exception + * block, is normally meant to be handled so the session carries on, + * unlike a top-level error, so a temporary slot is left in place. That + * matches the temporary slot behavior that predates this callback and is + * simpler to reason about; the slot lives on until the session ends or a + * top-level error occurs, as documented. + */ + ReplicationSlotRelease(); } /* @@ -1042,6 +1107,7 @@ ReplicationSlotDropAcquired(bool try_disable) /* slot isn't acquired anymore */ MyReplicationSlot = NULL; + MyReplicationSlotSubId = InvalidSubTransactionId; ReplicationSlotDropPtr(slot); diff --git a/src/backend/tcop/postgres.c b/src/backend/tcop/postgres.c index b6bdfe213fe..a2f72f9eb5b 100644 --- a/src/backend/tcop/postgres.c +++ b/src/backend/tcop/postgres.c @@ -4638,7 +4638,9 @@ PostgresMain(const char *dbname, const char *username) * need to be able to start and abort transactions while having a slot * acquired. But we never need to hold them across top level errors, * so releasing here is fine. There also is a before_shmem_exit() - * callback ensuring correct cleanup on FATAL errors. + * callback for FATAL errors, and AtEOSubXact_ReplicationSlot() for an + * error caught in a subtransaction. The latter only releases the slot + * and, unlike here, does not drop the session's temporary slots. */ if (MyReplicationSlot != NULL) ReplicationSlotRelease(); diff --git a/src/include/replication/slot.h b/src/include/replication/slot.h index 9b29444cbca..171ba6f2318 100644 --- a/src/include/replication/slot.h +++ b/src/include/replication/slot.h @@ -343,6 +343,7 @@ extern void ReplicationSlotAcquire(const char *name, bool nowait, bool error_if_invalid); extern void ReplicationSlotRelease(void); extern void ReplicationSlotCleanup(bool synced_only); +extern void AtEOSubXact_ReplicationSlot(bool isCommit, SubTransactionId mySubid); extern void ReplicationSlotSave(void); extern void ReplicationSlotMarkDirty(void); -- 2.47.3
