Commit 0bada39c83 fixed a bug of this kind,
which existed in all branches for six days before detection. While the
probability of reaching the trouble was low, the disruption was extreme. No
new backends could start, and service restoration needed an immediate
shutdown. Hence, add this to catch the next bug like it.
The new check in RelationIdGetRelation() suffices to make autovacuum detect
the bug in commit 243e9b40f1 that led to commit
0bada39. This also checks in a number of similar places. It replaces each
Assert(IsTransactionState()) that pertained to a conditional catalog read.
Back-patch to v14 - v17. This a back-patch of commit
f4ece891fc (from before v18 branched) to
all supported branches, to accompany the back-patch of commits 243e9b4
and 0bada39. For catalog indexes, the bttextcmp() behavior that
motivated IsCatalogTextUniqueIndexOid() was v18-specific. Hence, this
back-patch doesn't need that or its correction from commit
4a4ee0c2c1.
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/20250410191830.0e.nmisch@google.com
Discussion: https://postgr.es/m/10ec0bc3-5933-1189-6bb8-5dec4114558e@gmail.com
Backpatch-through: 14-17
A buffer lock won't stop a reader having already checked tuple
visibility. If a vac_update_datfrozenid() and then a crash happened
during inplace update of a relfrozenxid value, datfrozenxid could
overtake relfrozenxid. That could lead to "could not access status of
transaction" errors.
Back-patch to v14 - v17. This is a back-patch of commits:
- 8e7e672cda
(main change, on master, before v18 branched)
- 8180136652
(defect fix, on master, before v18 branched)
It reverses commit bc6bad8857, my revert
of the original back-patch.
In v14, this also back-patches the assertion removal from commit
7fcf2faf9c.
Discussion: https://postgr.es/m/20240620012908.92.nmisch@google.com
Backpatch-through: 14-17
The inplace update survives ROLLBACK. The inval didn't, so another
backend's DDL could then update the row without incorporating the
inplace update. In the test this fixes, a mix of CREATE INDEX and ALTER
TABLE resulted in a table with an index, yet relhasindex=f. That is a
source of index corruption.
Back-patch to v14 - v17. This is a back-patch of commits:
- 243e9b40f1
(main change, on master, before v18 branched)
- 0bada39c83
(defect fix, on master, before v18 branched)
- bae8ca82fd
(cosmetics from post-commit review, on REL_18_STABLE)
It reverses commit c1099dd745, my revert
of the original back-patch of 243e9b4.
This back-patch omits the non-comment heap_decode() changes. I find
those changes removed harmless code that was last necessary in v13. See
discussion thread for details. The back branches aren't the place to
remove such code.
Like the original back-patch, this doesn't change WAL, because these
branches use end-of-recovery SIResetAll(). All branches change the ABI
of extern function PrepareToInvalidateCacheTuple(). No PGXN extension
calls that, and there's no apparent use case in extensions. Expect
".abi-compliance-history" edits to follow.
Reviewed-by: Paul A Jungwirth <pj@illuminatedcomputing.com>
Reviewed-by: Surya Poondla <s_poondla@apple.com>
Reviewed-by: Ilyasov Ian <ianilyasov@outlook.com>
Reviewed-by: Nitin Motiani <nitinmotiani@google.com> (in earlier versions)
Reviewed-by: Andres Freund <andres@anarazel.de> (in earlier versions)
Discussion: https://postgr.es/m/20240523000548.58.nmisch@google.com
Backpatch-through: 14-17
This file separates public and static functions with a separator
comment, but two routines were not defined in a location reflecting
that, so reorder them.
Back-patch commit c2bdd2c5b1 to v15 - v16.
This avoids merge conflicts in the next commit, which modifies a
function this moved. Exclude v14, which is so different that the merge
conflict savings would be immaterial.
Author: Aleksander Alekseev <aleksander@timescale.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAJ7c6TMX2dd0g91UKvcC+CVygKQYJkKJq1+ZzT4rOK42+b53=w@mail.gmail.com
Backpatch-through: 15-16
Previously, the API for ltree_strncasecmp() took two inputs but only
one length (that of the smaller input). It truncated the larger input
to that length, but that could break a multibyte sequence.
Change the API to be a check for prefix equality (possibly
case-insensitive) instead, which is all that's needed by the
callers. Also, provide the lengths of both inputs.
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Peter Eisentraut <peter@eisentraut.org>
Discussion: https://postgr.es/m/5f65b85740197ba6249ea507cddf609f84a6188b.camel%40j-davis.com
Backpatch-through: 14
We already do this in CreateParallelContext, InitializeParallelDSM, and
LaunchParallelWorkers. I suspect the reason why the matching logic was
omitted from ReinitializeParallelDSM is that I failed to realize that
any memory allocation was happening here -- but shm_mq_attach does
allocate, which could result in a shm_mq_handle being allocated in a
shorter-lived context than the ParallelContext which points to it.
That could result in a crash if the shorter-lived context is freed
before the parallel context is destroyed. As far as I am currently
aware, there is no way to reach a crash using only code that is
present in core PostgreSQL, but extensions could potentially trip
over this. Fixing this in the back-branches appears low-risk, so
back-patch to all supported versions.
Author: Jakub Wartak <jakub.wartak@enterprisedb.com>
Co-authored-by: Jeevan Chalke <jeevan.chalke@enterprisedb.com>
Backpatch-through: 14
Discussion: http://postgr.es/m/CAKZiRmwfVripa3FGo06=5D1EddpsLu9JY2iJOTgbsxUQ339ogQ@mail.gmail.com
This commit adds an extra check at the beginning of recovery to ensure
that the redo record of a checkpoint exists before attempting WAL
replay, logging a PANIC if the redo record referenced by the checkpoint
record could not be found. This is the same level of failure as when a
checkpoint record is missing. This check is added when a cluster is
started without a backup_label, after retrieving its checkpoint record.
The redo LSN used for the check is retrieved from the checkpoint record
successfully read.
In the case where a backup_label exists, the startup process already
fails if the redo record cannot be found after reading a checkpoint
record at the beginning of recovery.
Previously, the presence of the redo record was not checked. If the
redo and checkpoint records were located on different WAL segments, it
would be possible to miss a entire range of WAL records that should have
been replayed but were just ignored. The consequences of missing the
redo record depend on the version dealt with, these becoming worse the
older the version used:
- On HEAD, v18 and v17, recovery fails with a pointer dereference at the
beginning of the redo loop, as the redo record is expected but cannot be
found. These versions are good students, because we detect a failure
before doing anything, even if the failure is misleading in the shape of
a segmentation fault, giving no information that the redo record is
missing.
- In v16 and v15, problems show at the end of recovery within
FinishWalRecovery(), the startup process using a buggy LSN to decide
from where to start writing WAL. The cluster gets corrupted, still it
is noisy about it.
- v14 and older versions are worse: a cluster gets corrupted but it is
entirely silent about the matter. The redo record missing causes the
startup process to skip entirely recovery, because a missing record is
the same as not redo being required at all. This leads to data loss, as
everything is missed between the redo record and the checkpoint record.
Note that I have tested that down to 9.4, reproducing the issue with a
version of the author's reproducer slightly modified. The code is wrong
since at least 9.2, but I did not look at the exact point of origin.
This problem has been found by debugging a cluster where the WAL segment
including the redo segment was missing due to an operator error, leading
to a crash, based on an investigation in v15.
Requesting archive recovery with the creation of a recovery.signal or
a standby.signal even without a backup_label would mitigate the issue:
if the record cannot be found in pg_wal/, the missing segment can be
retrieved with a restore_command when checking that the redo record
exists. This was already the case without this commit, where recovery
would re-fetch the WAL segment that includes the redo record. The check
introduced by this commit makes the segment to be retrieved earlier to
make sure that the redo record can be found.
On HEAD, the code will be slightly changed in a follow-up commit to not
rely on a PANIC, to include a test able to emulate the original problem.
This is a minimal backpatchable fix, kept separated for clarity.
Reported-by: Andres Freund <andres@anarazel.de>
Analyzed-by: Andres Freund <andres@anarazel.de>
Author: Nitin Jadhav <nitinjadhavpostgres@gmail.com>
Discussion: https://postgr.es/m/20231023232145.cmqe73stvivsmlhs@awork3.anarazel.de
Discussion: https://postgr.es/m/CAMm1aWaaJi2w49c0RiaDBfhdCL6ztbr9m=daGqiOuVdizYWYaA@mail.gmail.com
Backpatch-through: 14
Coverity complained that offset cannot be 0 here because there's an
explicit check for "offset == 0" earlier in the function, but it
didn't see the possibility that offset could've wrapped around to 0.
The code is correct, but clarify the comment about it.
The same code exists in backbranches in the server
GetMultiXactIdMembers() function and in 'master' in the pg_upgrade
GetOldMultiXactIdSingleMember function. In backbranches Coverity
didn't complain about it because the check was merely an assertion,
but change the comment in all supported branches for consistency.
Per Tom Lane's suggestion.
Discussion: https://www.postgresql.org/message-id/1827755.1765752936@sss.pgh.pa.us
An array of LLVMBasicBlockRef is allocated with the size used for an
element being "LLVMBasicBlockRef *" rather than "LLVMBasicBlockRef".
LLVMBasicBlockRef is a type that refers to a pointer, so this did not
directly cause a problem because both should have the same size, still
it is incorrect.
This issue has been spotted while reviewing a different patch, and
exists since 2a0faed9d7, so backpatch all the way down.
Discussion: https://postgr.es/m/CA+hUKGLngd9cKHtTUuUdEo2eWEgUcZ_EQRbP55MigV2t_zTReg@mail.gmail.com
Backpatch-through: 14
The description of deferrable constraints in create_table.sgml states
that deferrable constraints cannot be used as conflict arbitrators in
an INSERT with an ON CONFLICT DO UPDATE clause, but in fact this
restriction applies to all ON CONFLICT clauses, not just those with DO
UPDATE. Fix this, and while at it, change the word "arbitrators" to
"arbiters", to match the terminology used elsewhere.
Author: Dean Rasheed <dean.a.rasheed@gmail.com>
Discussion: https://postgr.es/m/CAEZATCWsybvZP3ce8rGcVNx-QHuDOJZDz8y=p1SzqHwjRXyV4Q@mail.gmail.com
Backpatch-through: 14
This commit adds the version information of a node initialized by
Cluster.pm, that may vary depending on the install_path given by the
test. The code was written so as the node information, that includes
the version number, was dumped before the version number was set.
This is particularly useful for the pg_upgrade TAP tests, that may mix
several versions for cross-version runs. The TAP infrastructure also
allows mixing nodes with different versions, so this information can be
useful for out-of-core tests.
Backpatch down to v15, where Cluster.pm and the pg_upgrade TAP tests
have been introduced.
Author: Potapov Alexander <a.potapov@postgrespro.com>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Discussion: https://postgr.es/m/e59bb-692c0a80-5-6f987180@170377126
Backpatch-through: 15
With this commit, the next multixid's offset will always be set on the
offsets page, by the time that a backend might try to read it, so we
no longer need the waiting mechanism with the condition variable. In
other words, this eliminates "corner case 2" mentioned in the
comments.
The waiting mechanism was broken in a few scenarios:
- When nextMulti was advanced without WAL-logging the next
multixid. For example, if a later multixid was already assigned and
WAL-logged before the previous one was WAL-logged, and then the
server crashed. In that case the next offset would never be set in
the offsets SLRU, and a query trying to read it would get stuck
waiting for it. Same thing could happen if pg_resetwal was used to
forcibly advance nextMulti.
- In hot standby mode, a deadlock could happen where one backend waits
for the next multixid assignment record, but WAL replay is not
advancing because of a recovery conflict with the waiting backend.
The old TAP test used carefully placed injection points to exercise
the old waiting code, but now that the waiting code is gone, much of
the old test is no longer relevant. Rewrite the test to reproduce the
IPC/MultixactCreation hang after crash recovery instead, and to verify
that previously recorded multixids stay readable.
Backpatch to all supported versions. In back-branches, we still need
to be able to read WAL that was generated before this fix, so in the
back-branches this includes a hack to initialize the next offsets page
when replaying XLOG_MULTIXACT_CREATE_ID for the last multixid on a
page. On 'master', bump XLOG_PAGE_MAGIC instead to indicate that the
WAL is not compatible.
Author: Andrey Borodin <amborodin@acm.org>
Reviewed-by: Dmitry Yurichev <dsy.075@yandex.ru>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Kirill Reshke <reshkekirill@gmail.com>
Reviewed-by: Ivan Bykov <i.bykov@modernsys.ru>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Discussion: https://www.postgresql.org/message-id/172e5723-d65f-4eec-b512-14beacb326ce@yandex.ru
Backpatch-through: 14
amcheck incorrectly reported the following error if there were any
half-dead pages in the index:
ERROR: mismatch between parent key and child high key in index
"amchecktest_id_idx"
It's expected that a half-dead page does not have a downlink in the
parent level, so skip the test.
Reported-by: Konstantin Knizhnik <knizhnik@garret.ru>
Reviewed-by: Peter Geoghegan <pg@bowt.ie>
Reviewed-by: Mihail Nikalayeu <mihailnikalayeu@gmail.com>
Discussion: https://www.postgresql.org/message-id/33e39552-6a2a-46f3-8b34-3f9f8004451f@garret.ru
Backpatch-through: 14
Formerly, when updating an auto-updatable view, or a relation with
rules, if the original query had any data-modifying CTEs, the rewriter
would rewrite those CTEs multiple times as RewriteQuery() recursed
into the product queries. In most cases that was harmless, because
RewriteQuery() is mostly idempotent. However, if the CTE involved
updating an always-generated column, it would trigger an error because
any subsequent rewrite would appear to be attempting to assign a
non-default value to the always-generated column.
This could perhaps be fixed by attempting to make RewriteQuery() fully
idempotent, but that looks quite tricky to achieve, and would probably
be quite fragile, given that more generated-column-type features might
be added in the future.
Instead, fix by arranging for RewriteQuery() to rewrite each CTE
exactly once (by tracking the number of CTEs already rewritten as it
recurses). This has the advantage of being simpler and more efficient,
but it does make RewriteQuery() dependent on the order in which
rewriteRuleAction() joins the CTE lists from the original query and
the rule action, so care must be taken if that is ever changed.
Reported-by: Bernice Southey <bernice.southey@gmail.com>
Author: Bernice Southey <bernice.southey@gmail.com>
Author: Dean Rasheed <dean.a.rasheed@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Kirill Reshke <reshkekirill@gmail.com>
Discussion: https://postgr.es/m/CAEDh4nyD6MSH9bROhsOsuTqGAv_QceU_GDvN9WcHLtZTCYM1kA@mail.gmail.com
Backpatch-through: 14
Normally, if a WHERE clause is implied by the predicate of a partial
index, we drop that clause from the set of quals used with the index,
since it's redundant to test it if we're scanning that index.
However, if it's a hash index (or any !amoptionalkey index), this
could result in dropping all available quals for the index's first
key, preventing us from generating an indexscan.
It's fair to question the practical usefulness of this case. Since
hash only supports equality quals, the situation could only arise
if the index's predicate is "WHERE indexkey = constant", implying
that the index contains only one hash value, which would make hash
a really poor choice of index type. However, perhaps there are
other !amoptionalkey index AMs out there with which such cases are
more plausible.
To fix, just don't filter the candidate indexquals this way if
the index is !amoptionalkey. That's a bit hokey because it may
result in testing quals we didn't need to test, but to do it
more accurately we'd have to redundantly identify which candidate
quals are actually usable with the index, something we don't know
at this early stage of planning. Doesn't seem worth the effort.
Reported-by: Sergei Glukhov <s.glukhov@postgrespro.ru>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/e200bf38-6b45-446a-83fd-48617211feff@postgrespro.ru
Backpatch-through: 14
The documentation for CREATE/ALTER PUBLICATION previously showed:
[ ONLY ] table_name [ * ] [ ( column_name [, ... ] ) ] [ WHERE ( expression ) ] [, ... ]
to indicate that the table/column specification could be repeated.
However, placing [, ... ] directly after a multi-part construct was
misleading and made it unclear which portion was repeatable.
This commit introduces a new term, table_and_columns, to represent:
[ ONLY ] table_name [ * ] [ ( column_name [, ... ] ) ] [ WHERE ( expression ) ]
and updates the synopsis to use:
table_and_columns [, ... ]
which clearly identifies the repeatable element.
Backpatched to v15, where the misleading syntax was introduced.
Author: Peter Smith <smithpb2250@gmail.com>
Reviewed-by: Chao Li <lic@highgo.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/CAHut+PtsyvYL3KmA6C8f0ZpXQ=7FEqQtETVy-BOF+cm9WPvfMQ@mail.gmail.com
Backpatch-through: 15
When running on Windows (or EXEC_BACKEND) the SSL configuration will
be reloaded on each backend start, so the passphrase command will be
reloaded along with it. This implies that passphrase command reload
must be enabled on Windows for connections to work at all. Document
this since it wasn't mentioned explicitly, and will there add markup
for parameter value to match the rest of the docs.
Backpatch to all supported versions.
Author: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Peter Eisentraut <peter@eisentraut.org>
Discussion: https://postgr.es/m/5F301096-921A-427D-8EC1-EBAEC2A35082@yesql.se
Backpatch-through: 14
Accidentally the code in LWLockWakeup() checked the list of to-be-woken up
processes to see if LW_FLAG_HAS_WAITERS should be unset. That means that
HAS_WAITERS would not get unset immediately, but only during the next,
unnecessary, call to LWLockWakeup().
Luckily, as the code stands, this is just a small efficiency issue.
However, if there were (as in a patch of mine) a case in which LWLockWakeup()
would not find any backend to wake, despite the wait list not being empty,
we'd wrongly unset LW_FLAG_HAS_WAITERS, leading to potentially hanging.
While the consequences in the backbranches are limited, the code as-is
confusing, and it is possible that there are workloads where the additional
wait list lock acquisitions hurt, therefore backpatch.
Discussion: https://postgr.es/m/fvfmkr5kk4nyex56ejgxj3uzi63isfxovp2biecb4bspbjrze7@az2pljabhnff
Backpatch-through: 14
The fix for bug #19055 (commit b0cc0a71e) allowed CTE references in
sub-selects within aggregate functions to affect the semantic levels
assigned to such aggregates. It turns out this broke some related
cases, leading to assertion failures or strange planner errors such
as "unexpected outer reference in CTE query". After experimenting
with some alternative rules for assigning the semantic level in
such cases, we've come to the conclusion that changing the level
is more likely to break things than be helpful.
Therefore, this patch undoes what b0cc0a71e changed, and instead
installs logic to throw an error if there is any reference to a
CTE that's below the semantic level that standard SQL rules would
assign to the aggregate based on its contained Var and Aggref nodes.
(The SQL standard disallows sub-selects within aggregate functions,
so it can't reach the troublesome case and hence has no rule for
what to do.)
Perhaps someone will come along with a legitimate query that this
logic rejects, and if so probably the example will help us craft
a level-adjustment rule that works better than what b0cc0a71e did.
I'm not holding my breath for that though, because the previous
logic had been there for a very long time before bug #19055 without
complaints, and that bug report sure looks to have originated from
fuzzing not from real usage.
Like b0cc0a71e, back-patch to all supported branches, though
sadly that no longer includes v13.
Bug: #19106
Reported-by: Kamil Monicz <kamil@monicz.dev>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/19106-9dd3668a0734cd72@postgresql.org
Backpatch-through: 14
As noted in the commit message for 5e4fcbe531, the addition of a
second parameter to CreateStatistics() breaks ABI compatibility,
but we are unaware of any impacted third-party code. This commit
updates .abi-compliance-history accordingly.
Backpatch-through: 14-18
Commit 5e4fcbe531 added a check_rights parameter to this function
for use by ALTER TABLE commands that re-create statistics objects.
However, we intentionally ignore check_rights when verifying
relation ownership because this function's lookup could return a
different answer than the caller's. This commit adds a note to
this effect so that we remember it down the road.
Reviewed-by: Noah Misch <noah@leadboat.com>
Backpatch-through: 14
On the CREATE POLICY page, the "Policies Applied by Command Type"
table was missing MERGE ... THEN DELETE and some of the policies
applied during INSERT ... ON CONFLICT and MERGE. Fix that, and try to
improve readability by listing the various MERGE cases separately,
rather than together with INSERT/UPDATE/DELETE. Mention COPY ... TO
along with SELECT, since it behaves in the same way. In addition,
document which policy violations cause errors to be thrown, and which
just cause rows to be silently ignored.
Also, a paragraph above the table states that INSERT ... ON CONFLICT
DO UPDATE only checks the WITH CHECK expressions of INSERT policies
for rows appended to the relation by the INSERT path, which is
incorrect -- all rows proposed for insertion are checked, regardless
of whether they end up being inserted. Fix that, and also mention that
the same applies to INSERT ... ON CONFLICT DO NOTHING.
In addition, in various other places on that page, clarify how the
different types of policy are applied to different commands, and
whether or not errors are thrown when policy checks do not pass.
Backpatch to all supported versions. Prior to v17, MERGE did not
support RETURNING, and so MERGE ... THEN INSERT would never check new
rows against SELECT policies. Prior to v15, MERGE was not supported at
all.
Author: Dean Rasheed <dean.a.rasheed@gmail.com>
Reviewed-by: Viktor Holmberg <v@viktorh.net>
Reviewed-by: Jian He <jian.universality@gmail.com>
Discussion: https://postgr.es/m/CAEZATCWqnfeChjK=n1V_dYZT4rt4mnq+ybf9c0qXDYTVMsy8pg@mail.gmail.com
Backpatch-through: 14
Before we started to freeze async notify entries (commit 8eeb4a0f7c),
no one looked at the 'xid' on an entry with invalid 'dboid'. But now
we might actually need to freeze it later. Initialize them with
InvalidTransactionId to begin with, to avoid that work later.
Álvaro pointed this out in review of commit 8eeb4a0f7c, but I forgot
to include this change there.
Author: Álvaro Herrera <alvherre@kurilemu.de>
Discussion: https://www.postgresql.org/message-id/202511071410.52ll56eyixx7@alvherre.pgsql
Backpatch-through: 14
Previous commit fixed a bug where VACUUM would truncate the CLOG
that's still needed to check the commit status of XIDs in the async
notify queue, but as mentioned in the commit message, it wasn't a full
fix. If a backend is executing asyncQueueReadAllNotifications() and
has just made a local copy of an async SLRU page which contains old
XIDs, vacuum can concurrently truncate the CLOG covering those XIDs,
and the backend still gets an error when it calls
TransactionIdDidCommit() on those XIDs in the local copy. This commit
fixes that race condition.
To fix, hold the SLRU bank lock across the TransactionIdDidCommit()
calls in NOTIFY processing.
Per Tom Lane's idea. Backpatch to all supported versions.
Reviewed-by: Joel Jacobson <joel@compiler.org>
Reviewed-by: Arseniy Mukhin <arseniy.mukhin.dev@gmail.com>
Discussion: https://www.postgresql.org/message-id/2759499.1761756503@sss.pgh.pa.us
Backpatch-through: 14
The async notification queue contains the XID of the sender, and when
processing notifications we call TransactionIdDidCommit() on the
XID. But we had no safeguards to prevent the CLOG segments containing
those XIDs from being truncated away. As a result, if a backend didn't
for some reason process its notifications for a long time, or when a
new backend issued LISTEN, you could get an error like:
test=# listen c21;
ERROR: 58P01: could not access status of transaction 14279685
DETAIL: Could not open file "pg_xact/000D": No such file or directory.
LOCATION: SlruReportIOError, slru.c:1087
To fix, make VACUUM "freeze" the XIDs in the async notification queue
before truncating the CLOG. Old XIDs are replaced with
FrozenTransactionId or InvalidTransactionId.
Note: This commit is not a full fix. A race condition remains, where a
backend is executing asyncQueueReadAllNotifications() and has just
made a local copy of an async SLRU page which contains old XIDs, while
vacuum concurrently truncates the CLOG covering those XIDs. When the
backend then calls TransactionIdDidCommit() on those XIDs from the
local copy, you still get the error. The next commit will fix that
remaining race condition.
This was first reported by Sergey Zhuravlev in 2021, with many other
people hitting the same issue later. Thanks to:
- Alexandra Wang, Daniil Davydov, Andrei Varashen and Jacques Combrink
for investigating and providing reproducable test cases,
- Matheus Alcantara and Arseniy Mukhin for review and earlier proposed
patches to fix this,
- Álvaro Herrera and Masahiko Sawada for reviews,
- Yura Sokolov aka funny-falcon for the idea of marking transactions
as committed in the notification queue, and
- Joel Jacobson for the final patch version. I hope I didn't forget
anyone.
Backpatch to all supported versions. I believe the bug goes back all
the way to commit d1e027221d, which introduced the SLRU-based async
notification queue.
Discussion: https://www.postgresql.org/message-id/16961-25f29f95b3604a8a@postgresql.org
Discussion: https://www.postgresql.org/message-id/18804-bccbbde5e77a68c2@postgresql.org
Discussion: https://www.postgresql.org/message-id/CAK98qZ3wZLE-RZJN_Y%2BTFjiTRPPFPBwNBpBi5K5CU8hUHkzDpw@mail.gmail.com
Backpatch-through: 14
Previously, if async notify processing encountered an error, we would
report the error to the client and advance our read position past the
offending entry to prevent trying to process it over and over
again. Trying to continue after an error has a few problems however:
- We have no way of telling the client that a notification was
lost. They get an ERROR, but that doesn't tell you much. As such,
it's not clear if keeping the connection alive after losing a
notification is a good thing. Depending on the application logic,
missing a notification could cause the application to get stuck
waiting, for example.
- If the connection is idle, PqCommReadingMsg is set and any ERROR is
turned into FATAL anyway.
- We bailed out of the notification processing loop on first error
without processing any subsequent notifications. The subsequent
notifications would not be processed until another notify interrupt
arrives. For example, if there were two notifications pending, and
processing the first one caused an ERROR, the second notification
would not be processed until someone sent a new NOTIFY.
This commit changes the behavior so that any ERROR while processing
async notifications is turned into FATAL, causing the client
connection to be terminated. That makes the behavior more consistent
as that's what happened in idle state already, and terminating the
connection is a clear signal to the application that it might've
missed some notifications.
The reason to do this now is that the next commits will change the
notification processing code in a way that would make it harder to
skip over just the offending notification entry on error.
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Arseniy Mukhin <arseniy.mukhin.dev@gmail.com>
Discussion: https://www.postgresql.org/message-id/fedbd908-4571-4bbe-b48e-63bfdcc38f64@iki.fi
Backpatch-through: 14
The synopsis for the ALTER PUBLICATION ... DROP ... command incorrectly
implied that a column list and WHERE clause could be specified as part of
the publication object. However, these options are not allowed for
DROP operations, making the documentation misleading.
This commit corrects the synopsis to clearly show only the valid forms
of publication objects.
Backpatched to v15, where the incorrect synopsis was introduced.
Author: Peter Smith <smithpb2250@gmail.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/CAHut+PsPu+47Q7b0o6h1r-qSt90U3zgbAHMHUag5o5E1Lo+=uw@mail.gmail.com
Backpatch-through: 15
This omission allowed table owners to create statistics in any
schema, potentially leading to unexpected naming conflicts. For
ALTER TABLE commands that require re-creating statistics objects,
skip this check in case the user has since lost CREATE on the
schema. The addition of a second parameter to CreateStatistics()
breaks ABI compatibility, but we are unaware of any impacted
third-party code.
Reported-by: Jelte Fennema-Nio <postgres@jeltef.nl>
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Co-authored-by: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Security: CVE-2025-12817
Backpatch-through: 13
Several functions could overflow their size calculations, when presented
with very large inputs from remote and/or untrusted locations, and then
allocate buffers that were too small to hold the intended contents.
Switch from int to size_t where appropriate, and check for overflow
conditions when the inputs could have plausibly originated outside of
the libpq trust boundary. (Overflows from within the trust boundary are
still possible, but these will be fixed separately.) A version of
add_size() is ported from the backend to assist with code that performs
more complicated concatenation.
Reported-by: Aleksey Solovev (Positive Technologies)
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Security: CVE-2025-12818
Backpatch-through: 13
generic-gcc.h maps our read and write barriers to C11 acquire and
release fences using compiler builtins, for platforms where we don't
have our own hand-rolled assembler. This is apparently enough for GCC,
but the C11 memory model is only defined in terms of atomic accesses,
and our barriers for non-atomic, non-volatile accesses were not always
respected under Clang's stricter interpretation of the standard.
This explains the occasional breakage observed on new RISC-V + Clang
animal greenfly in lock-free PgAioHandle manipulation code containing a
repeating pattern of loads and read barriers. The problem can also be
observed in code generated for MIPS and LoongAarch, though we aren't
currently testing those with Clang, and on x86, though we use our own
assembler there. The scariest aspect is that we use the generic version
on very common ARM systems, but it doesn't seem to reorder the relevant
code there (or we'd have debugged this long ago).
Fix by inserting an explicit compiler barrier. It expands to an empty
assembler block declared to have memory side-effects, so registers are
flushed and reordering is prevented. In those respects this is like the
architecture-specific assembler versions, but the compiler is still in
charge of generating the appropriate fence instruction. Done for write
barriers on principle, though concrete problems have only been observed
with read barriers.
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Tested-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/d79691be-22bd-457d-9d90-18033b78c40a%40gmail.com
Backpatch-through: 13
The following parameters can only be set at server start because
their context is PGC_POSTMASTER, but this information was missing
or incorrectly documented. This commit adds or corrects
that information for the following parameters:
* debug_io_direct
* dynamic_shared_memory_type
* event_source
* huge_pages
* io_max_combine_limit
* max_notify_queue_pages
* shared_memory_type
* track_commit_timestamp
* wal_decode_buffer_size
Backpatched to all supported branches.
Author: Karina Litskevich <litskevichkarina@gmail.com>
Reviewed-by: Chao Li <lic@highgo.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/CAHGQGwGfPzcin-_6XwPgVbWTOUFVZgHF5g9ROrwLUdCTfjy=0A@mail.gmail.com
Backpatch-through: 13
XLogRecPtrIsInvalid() is inconsistent with the affirmative form of
macros used for other datatypes, and leads to awkward double negatives
in a few places. This commit introduces XLogRecPtrIsValid(), which
allows code to be written more naturally.
This patch only adds the new macro. XLogRecPtrIsInvalid() is left in
place, and all existing callers remain untouched. This means all
supported branches can accept hypothetical bug fixes that use the new
macro, and at the same time any code that compiled with the original
formulation will continue to silently compile just fine.
Author: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Backpatch-through: 13
Discussion: https://postgr.es/m/aQB7EvGqrbZXrMlg@ip-10-97-1-34.eu-west-3.compute.internal
