While the preceding commit prevented such attachments from occurring
in future, this one aims to prevent further abuse of any already-
created operator that exposes _int_matchsel to the wrong data types.
(No other contrib module has a vulnerable selectivity estimator.)
We need only check that the Const we've found in the query is indeed
of the type we expect (query_int), but there's a difficulty: as an
extension type, query_int doesn't have a fixed OID that we could
hard-code into the estimator.
Therefore, the bulk of this patch consists of infrastructure to let
an extension function securely look up the OID of a datatype
belonging to the same extension. (Extension authors have requested
such functionality before, so we anticipate that this code will
have additional non-security uses, and may soon be extended to allow
looking up other kinds of SQL objects.)
This is done by first finding the extension that owns the calling
function (there can be only one), and then thumbing through the
objects owned by that extension to find a type that has the desired
name. This is relatively expensive, especially for large extensions,
so a simple cache is put in front of these lookups.
Reported-by: Daniel Firer as part of zeroday.cloud
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Noah Misch <noah@leadboat.com>
Security: CVE-2026-2004
Backpatch-through: 14
Selectivity estimators come in two flavors: those that make specific
assumptions about the data types they are working with, and those
that don't. Most of the built-in estimators are of the latter kind
and are meant to be safely attachable to any operator. If the
operator does not behave as the estimator expects, you might get a
poor estimate, but it won't crash.
However, estimators that do make datatype assumptions can malfunction
if they are attached to the wrong operator, since then the data they
get from pg_statistic may not be of the type they expect. This can
rise to the level of a security problem, even permitting arbitrary
code execution by a user who has the ability to create SQL objects.
To close this hole, establish a rule that built-in estimators are
required to protect themselves against being called on the wrong type
of data. It does not seem practical however to expect estimators in
extensions to reach a similar level of security, at least not in the
near term. Therefore, also establish a rule that superuser privilege
is required to attach a non-built-in estimator to an operator.
We expect that this restriction will have little negative impact on
extensions, since estimators generally have to be written in C and
thus superuser privilege is required to create them in the first
place.
This commit changes the privilege checks in CREATE/ALTER OPERATOR
to enforce the rule about superuser privilege, and fixes a couple
of built-in estimators that were making datatype assumptions without
sufficiently checking that they're valid.
Reported-by: Daniel Firer as part of zeroday.cloud
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Noah Misch <noah@leadboat.com>
Security: CVE-2026-2004
Backpatch-through: 14
An upcoming patch requires this cache so that it can track updates
in the pg_extension catalog. So far though, the EXTENSIONOID cache
only exists in v18 and up (see 490f869d9). We can add it in older
branches without an ABI break, if we are careful not to disturb the
numbering of existing syscache IDs.
In v16 and before, that just requires adding the new ID at the end
of the hand-assigned enum list, ignoring our convention about
alphabetizing the IDs. But in v17, genbki.pl enforces alphabetical
order of the IDs listed in MAKE_SYSCACHE macros. We can fake it
out by calling the new cache ZEXTENSIONOID.
Note that adding a syscache does change the required contents of the
relcache init file (pg_internal.init). But that isn't problematic
since we blow those away at postmaster start for other reasons.
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Noah Misch <noah@leadboat.com>
Security: CVE-2026-2004
Backpatch-through: 14-17
These data types are represented like full-fledged arrays, but
functions that deal specifically with these types assume that the
array is 1-dimensional and contains no nulls. However, there are
cast pathways that allow general oid[] or int2[] arrays to be cast
to these types, allowing these expectations to be violated. This
can be exploited to cause server memory disclosure or SIGSEGV.
Fix by installing explicit checks in functions that accept these
types.
Reported-by: Altan Birler <altan.birler@tum.de>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Noah Misch <noah@leadboat.com>
Security: CVE-2026-2003
Backpatch-through: 14
pgp_sym_decrypt() and pgp_pub_decrypt() will raise such errors, while
bytea variants will not. The existing "dat3" test decrypted to non-UTF8
text, so switch that query to bytea.
The long-term intent is for type "text" to always be valid in the
database encoding. pgcrypto has long been known as a source of
exceptions to that intent, but a report about exploiting invalid values
of type "text" brought this module to the forefront. This particular
exception is straightforward to fix, with reasonable effect on user
queries. Back-patch to v14 (all supported versions).
Reported-by: Paul Gerste (as part of zeroday.cloud)
Reported-by: Moritz Sanft (as part of zeroday.cloud)
Author: shihao zhong <zhong950419@gmail.com>
Reviewed-by: cary huang <hcary328@gmail.com>
Discussion: https://postgr.es/m/CAGRkXqRZyo0gLxPJqUsDqtWYBbgM14betsHiLRPj9mo2=z9VvA@mail.gmail.com
Backpatch-through: 14
Security: CVE-2026-2006
A security patch changed them today, so close the coverage gap now.
Test that buffer overrun is avoided when pg_mblen*() requires more
than the number of bytes remaining.
This does not cover the calls in dict_thesaurus.c or in dict_synonym.c.
That code is straightforward. To change that code's input, one must
have access to modify installed OS files, so low-privilege users are not
a threat. Testing this would likewise require changing installed
share/postgresql/tsearch_data, which was enough of an obstacle to not
bother.
Security: CVE-2026-2006
Backpatch-through: 14
Co-authored-by: Thomas Munro <thomas.munro@gmail.com>
Co-authored-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
A corrupted string could cause code that iterates with pg_mblen() to
overrun its buffer. Fix, by converting all callers to one of the
following:
1. Callers with a null-terminated string now use pg_mblen_cstr(), which
raises an "illegal byte sequence" error if it finds a terminator in the
middle of the sequence.
2. Callers with a length or end pointer now use either
pg_mblen_with_len() or pg_mblen_range(), for the same effect, depending
on which of the two seems more convenient at each site.
3. A small number of cases pre-validate a string, and can use
pg_mblen_unbounded().
The traditional pg_mblen() function and COPYCHAR macro still exist for
backward compatibility, but are no longer used by core code and are
hereby deprecated. The same applies to the t_isXXX() functions.
Security: CVE-2026-2006
Backpatch-through: 14
Co-authored-by: Thomas Munro <thomas.munro@gmail.com>
Co-authored-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Reported-by: Paul Gerste (as part of zeroday.cloud)
Reported-by: Moritz Sanft (as part of zeroday.cloud)
When converting multibyte to pg_wchar, the UTF-8 implementation would
silently ignore an incomplete final character, while the other
implementations would cast a single byte to pg_wchar, and then repeat
for the remaining byte sequence. While it didn't overrun the buffer, it
was surely garbage output.
Make all encodings behave like the UTF-8 implementation. A later change
for master only will convert this to an error, but we choose not to
back-patch that behavior change on the off-chance that someone is
relying on the existing UTF-8 behavior.
Security: CVE-2026-2006
Backpatch-through: 14
Author: Thomas Munro <thomas.munro@gmail.com>
Reported-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
While EUC_CN supports only 1- and 2-byte sequences (CS0, CS1), the
mb<->wchar conversion functions allow 3-byte sequences beginning SS2,
SS3.
Change pg_encoding_max_length() to return 3, not 2, to close a
hypothesized buffer overrun if a corrupted string is converted to wchar
and back again in a newly allocated buffer. We might reconsider that in
master (ie harmonizing in a different direction), but this change seems
better for the back-branches.
Also change pg_euccn_mblen() to report SS2 and SS3 characters as having
length 3 (following the example of EUC_KR). Even though such characters
would not pass verification, it's remotely possible that invalid bytes
could be used to compute a buffer size for use in wchar conversion.
Security: CVE-2026-2006
Backpatch-through: 14
Author: Thomas Munro <thomas.munro@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
pgp_pub_decrypt_bytea() was missing a safeguard for the session key
length read from the message data, that can be given in input of
pgp_pub_decrypt_bytea(). This can result in the possibility of a buffer
overflow for the session key data, when the length specified is longer
than PGP_MAX_KEY, which is the maximum size of the buffer where the
session data is copied to.
A script able to rebuild the message and key data that can trigger the
overflow is included in this commit, based on some contents provided by
the reporter, heavily editted by me. A SQL test is added, based on the
data generated by the script.
Reported-by: Team Xint Code as part of zeroday.cloud
Author: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Noah Misch <noah@leadboat.com>
Security: CVE-2026-2005
Backpatch-through: 14
This thinko caused us to not substitute our own getopt() code,
which results in failing to parse long options for the postmaster
since Solaris' getopt() doesn't do what we expect. This can be seen
in the results of buildfarm member icarus, which is the only one
trying to build via meson on Solaris.
Per consultation with pgsql-release, it seems okay to fix this
now even though we're in release freeze. The fix visibly won't
affect any other platforms, and it can't break Solaris/meson
builds any worse than they're already broken.
Discussion: https://postgr.es/m/2471229.1770499291@sss.pgh.pa.us
Backpatch-through: 16
It's not really an ABI break if you change the layout/size of an object
with incomplete type, as commit f94e9141 did, so advance the ABI
compliance reference commit in 16-18 to satisfy build farm animal crake.
Backpatch-through: 16-18
Discussion: https://www.postgresql.org/message-id/1871492.1770409863%40sss.pgh.pa.us
(This is a cherry-pick of 390b3cbbb, which I hadn't realized wasn't
backpatched. It was originally reported to security@ and determined not
to be a vulnerability; thanks to Stanislav Osipov for noticing the
omission in the back branches.)
In case of torn UTF8 in the input data we might end up going
past the end of the string since we don't account for length.
While validation won't be performed on a sequence with a NULL
byte it's better to avoid going past the end to beging with.
Fix by taking the length into consideration.
Reported-by: Stanislav Osipov <stasos24@gmail.com>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Discussion: https://postgr.es/m/CAOYmi+mTnmM172g=_+Yvc47hzzeAsYPy2C4UBY3HK9p-AXNV0g@mail.gmail.com
Backpatch-through: 14
These errors are very unlikely going to show up, but in the event that
they happen, some incorrect information would have been provided:
- In pg_rewind, a stat() failure was reported as an open() failure.
- In pg_combinebackup, a check for the new directory of a tablespace
mapping was referred as the old directory.
- In pg_combinebackup, a failure in reading a source file when copying
blocks referred to the destination file.
The changes for pg_combinebackup affect v17 and newer versions. For
pg_rewind, all the stable branches are affected.
Author: Man Zeng <zengman@halodbtech.com>
Discussion: https://postgr.es/m/tencent_1EE1430B1E6C18A663B8990F@qq.com
Backpatch-through: 14
Provide a way to disable the use of posix_fallocate() for relation
files. It was introduced by commit 4d330a61bb1. The new setting
file_extend_method=write_zeros can be used as a workaround for problems
reported from the field:
* BTRFS compression is disabled by the use of posix_fallocate()
* XFS could produce spurious ENOSPC errors in some Linux kernel
versions, though that problem is reported to have been fixed
The default is file_extend_method=posix_fallocate if available, as
before. The write_zeros option is similar to PostgreSQL < 16, except
that now it's multi-block.
Backpatch-through: 16
Reviewed-by: Jakub Wartak <jakub.wartak@enterprisedb.com>
Reported-by: Dimitrios Apostolou <jimis@gmx.net>
Discussion: https://postgr.es/m/b1843124-fd22-e279-a31f-252dffb6fbf2%40gmx.net
Commit 5f13999aa11 added a TAP test for GUC settings passed via the
CONNECTION string in logical replication, but the buildfarm member
sungazer reported test failures.
The test incorrectly used the subscriber's log file position as the
starting offset when reading the publisher's log. As a result, the test
failed to find the expected log message in the publisher's log and
erroneously reported a failure.
This commit fixes the test to use the publisher's own log file position
when reading the publisher's log.
Also, to avoid similar confusion in the future, this commit splits the single
$log_location variable into $log_location_pub and $log_location_sub,
clearly distinguishing publisher and subscriber log positions.
Backpatched to v15, where commit 5f13999aa11 introduced the test.
Per buildfarm member sungazer.
This issue was reported and diagnosed by Alexander Lakhin.
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/966ec3d8-1b6f-4f57-ae59-fc7d55bc9a5a@gmail.com
Backpatch-through: 15
Mostly this involves checking for NULL pointer before doing operations
that add a non-zero offset.
The exception is an overflow warning in heap_fetch_toast_slice(). This
was caused by unneeded parentheses forcing an expression to be
evaluated to a negative integer, which then got cast to size_t.
Per clang 21 undefined behavior sanitizer.
Backpatch to all supported versions.
Co-authored-by: Alexander Lakhin <exclusion@gmail.com>
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/777bd201-6e3a-4da0-a922-4ea9de46a3ee@gmail.com
Backpatch-through: 14
BackgroundPsql needs to wait for all the output from an interactive
psql command to come back. To make sure that's happened, it issues
the command, then issues \echo and \warn psql commands that echo
a "banner" string (which we assume won't appear in the command's
output), then waits for the banner strings to appear. The hazard
in this approach is that the banner will also appear in the echoed
psql commands themselves, so we need to distinguish those echoes from
the desired output. Commit 8b886a4e3 tried to do that by positing
that the desired output would be directly preceded and followed by
newlines, but it turns out that that assumption is timing-sensitive.
In particular, it tends to fail in builds made --without-readline,
wherein the command echoes will be made by the pty driver and may
be interspersed with prompts issued by psql proper.
It does seem safe to assume that the banner output we want will be
followed by a newline, since that should be the last output before
things quiesce. Therefore, we can improve matters by putting quotes
around the banner strings in the \echo and \warn psql commands, so
that their echoes cannot include banner directly followed by newline,
and then checking for just banner-and-newline in the match pattern.
While at it, spruce up the pump() call in sub query() to look like
the neater version in wait_connect(), and don't die on timeout
until after printing whatever we got.
Reported-by: Oleg Tselebrovskiy <o.tselebrovskiy@postgrespro.ru>
Diagnosed-by: Oleg Tselebrovskiy <o.tselebrovskiy@postgrespro.ru>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Soumya S Murali <soumyamurali.work@gmail.com>
Discussion: https://postgr.es/m/db6fdb35a8665ad3c18be01181d44b31@postgrespro.ru
Backpatch-through: 14
As noted in the commit message for b4307ae2e54, the change to the
TransitionCaptureState structure is nominally an ABI break, but it is
not expected to affect any third-party code. Therefore, add it to the
.abi-compliance-history file.
Discussion: https://postgr.es/m/19380-4e293be2b4007248%40postgresql.org
Backpatch-through: 15-18
ed1a88dda made it so WindowClauses can be merged when all window
functions belonging to the WindowClause can equally well use some
other WindowClause without any behavioral changes. When that
optimization applies, the WindowFunc's "winref" gets adjusted to
reference the new WindowClause.
That commit does not work well with the deduplication logic in
find_window_functions(), which only added the WindowFunc to the list
when there wasn't already an identical WindowFunc in the list. That
deduplication logic meant that the duplicate WindowFunc wouldn't get the
"winref" changed when optimize_window_clauses() was able to swap the
WindowFunc to another WindowClause. This could lead to the following
error in the unlikely event that the deduplication code did something and
the duplicate WindowFunc happened to be moved into another WindowClause.
ERROR: WindowFunc with winref 2 assigned to WindowAgg with winref 1
As it turns out, the deduplication logic in find_window_functions() is
pretty bogus. It might have done something when added, as that code
predates b8d7f053c, which changed how projections work. As it turns
out, at least now we *will* evaluate the duplicate WindowFuncs. All
that the deduplication code seems to do today is assist in
underestimating the WindowAggPath costs due to not counting the
evaluation costs of duplicate WindowFuncs.
Ideally the fix would be to remove the deduplication code, but that
could result in changes to the plan costs, as duplicate WindowFuncs
would then be costed. Instead, let's play it safe and shift the
deduplication code so it runs after the other processing in
optimize_window_clauses().
Backpatch only as far as v16 as there doesn't seem to be any other harm
done by the WindowFunc deduplication code before then. This issue was
fixed in master by 7027dd499.
Reported-by: Meng Zhang <mza117jc@gmail.com>
Author: Meng Zhang <mza117jc@gmail.com>
Author: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CAErYLFAuxmW0UVdgrz7iiuNrxGQnFK_OP9hBD5CUzRgjrVrz=Q@mail.gmail.com
Backpatch-through: 16
When executing a data-modifying CTE query containing MERGE and some
other DML operation on a table with statement-level AFTER triggers,
the transition tables passed to the triggers would fail to include the
rows affected by the MERGE.
The reason is that, when initializing a ModifyTable node for MERGE,
MakeTransitionCaptureState() would create a TransitionCaptureState
structure with a single "tcs_private" field pointing to an
AfterTriggersTableData structure with cmdType == CMD_MERGE. Tuples
captured there would then not be included in the sets of tuples
captured when executing INSERT/UPDATE/DELETE ModifyTable nodes in the
same query.
Since there are no MERGE triggers, we should only create
AfterTriggersTableData structures for INSERT/UPDATE/DELETE. Individual
MERGE actions should then use those, thereby sharing the same capture
tuplestores as any other DML commands executed in the same query.
This requires changing the TransitionCaptureState structure, replacing
"tcs_private" with 3 separate pointers to AfterTriggersTableData
structures, one for each of INSERT, UPDATE, and DELETE. Nominally,
this is an ABI break to a public structure in commands/trigger.h.
However, since this is a private field pointing to an opaque data
structure, the only way to create a valid TransitionCaptureState is by
calling MakeTransitionCaptureState(), and no extensions appear to be
doing that anyway, so it seems safe for back-patching.
Backpatch to v15, where MERGE was introduced.
Bug: #19380
Reported-by: Daniel Woelfel <dwwoelfel@gmail.com>
Author: Dean Rasheed <dean.a.rasheed@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/19380-4e293be2b4007248%40postgresql.org
Backpatch-through: 15
ExecInitModifyTable() unconditionally required a ctid junk column even
when the target was a partitioned table. This led to spurious "could
not find junk ctid column" errors when all children were excluded and
only the dummy root result relation remained.
A partitioned table only appears in the result relations list when all
leaf partitions have been pruned, leaving the dummy root as the sole
entry. Assert this invariant (nrels == 1) and skip the ctid requirement.
Also adjust ExecModifyTable() to tolerate invalid ri_RowIdAttNo for
partitioned tables, which is safe since no rows will be processed in
this case.
Bug: #19099
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Author: Amit Langote <amitlangote09@gmail.com>
Reviewed-by: Tender Wang <tndrwang@gmail.com>
Reviewed-by: Kirill Reshke <reshkekirill@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/19099-e05dcfa022fe553d%40postgresql.org
Backpatch-through: 14
Commit f16241bef mistakenly supposed that INSERT...ON CONFLICT DO
UPDATE rejects partitioned target tables. (This may have been
accurate when the patch was written, but it was already obsolete
when committed.) Hence, there's an assertion that we can't see
ItemPointerIndicatesMovedPartitions() in that path, but the assertion
is triggerable.
Some other places throw error if they see a moved-across-partitions
tuple, but there seems no need for that here, because if we just retry
then we get the same behavior as in the update-within-partition case,
as demonstrated by the new isolation test. So fix by deleting the
faulty Assert. (The fact that this is the fix doubtless explains
why we've heard no field complaints: the behavior of a non-assert
build is fine.)
The TM_Deleted case contains a cargo-culted copy of the same Assert,
which I also deleted to avoid confusion, although I believe that one
is actually not triggerable.
Per our code coverage report, neither the TM_Updated nor the
TM_Deleted case were reached at all by existing tests, so this
patch adds tests for both.
Reported-by: Dmitry Koval <d.koval@postgrespro.ru>
Author: Joseph Koshakow <koshy44@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/f5fffe4b-11b2-4557-a864-3587ff9b4c36@postgrespro.ru
Backpatch-through: 14
With LLVM >= 17, transform passes are provided as a string to
LLVMRunPasses. Only two strings were used: "default<O3>" and
"default<O0>,mem2reg".
With previous LLVM versions, an additional inline pass was added when
JIT inlining was enabled without optimization. With LLVM >= 17, the code
would go through llvm_inline, prepare the functions for inlining, but
the generated bitcode would be the same due to the missing inline pass.
This patch restores the previous behavior by adding an inline pass when
inlining is enabled but no optimization is done.
This fixes an oversight introduced by 76200e5e when support for LLVM 17
was added.
Backpatch-through: 14
Author: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Pierre Ducroquet <p.psql@pinaraf.info>
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Discussion: https://postgr.es/m/CAO6_XqrNjJnbn15ctPv7o4yEAT9fWa-dK15RSyun6QNw9YDtKg%40mail.gmail.com
We were using SnapshotAny to do some index checks, but that's wrong and
causes spurious errors when used on indexes created by CREATE INDEX
CONCURRENTLY. Fix it to use an MVCC snapshot, and add a test for it.
Backpatch of 6bd469d26aca to branches 14-16. I previously misidentified
the bug's origin: it came in with commit 7f563c09f890 (pg11-era, not
5ae2087202af as claimed previously), so all live branches are affected.
Also take the opportunity to fix some comments that we failed to update
in the original commits and apply pgperltidy. In branch 14, remove the
unnecessary test plan specification (which would have need to have been
changed anyway; c.f. commit 549ec201d613.)
Diagnosed-by: Donghang Lin <donghanglin@gmail.com>
Author: Mihail Nikalayeu <mihailnikalayeu@gmail.com>
Reviewed-by: Andrey Borodin <x4mmm@yandex-team.ru>
Backpatch-through: 17
Discussion: https://postgr.es/m/CANtu0ojmVd27fEhfpST7RG2KZvwkX=dMyKUqg0KM87FkOSdz8Q@mail.gmail.com
This is pretty pro-forma for our purposes, as the only change
is a historical correction for pre-1976 DST laws in
Baja California. (Upstream made this release mostly to update
their leap-second data, which we don't use.) But with minor
releases coming up, we should be up-to-date.
Backpatch-through: 14
The code adding the WAL information included in a backup manifest is
cross-checked with the contents of the timeline history file of the end
timeline. A check based on the end timeline, when it fails, reported
the value of the start timeline in the error message. This error is
fixed to show the correct timeline number in the report.
This error report would be confusing for users if seen, because it would
provide an incorrect information, so backpatch all the way down.
Oversight in 0d8c9c1210c4.
Author: Man Zeng <zengman@halodbtech.com>
Discussion: https://postgr.es/m/tencent_0F2949C4594556F672CF4658@qq.com
Backpatch-through: 14
If a FATAL error occurs while holding a lock in a DSM segment (such
as a dshash lock) and the process is not in a transaction, a
segmentation fault can occur during process exit.
The problem sequence is:
1. Process acquires a lock in a DSM segment (e.g., via dshash)
2. FATAL error occurs outside transaction context
3. proc_exit() begins, calling before_shmem_exit callbacks
4. dsm_backend_shutdown() detaches all DSM segments
5. Later, on_shmem_exit callbacks run
6. ProcKill() calls LWLockReleaseAll()
7. Segfault: the lock being released is in unmapped memory
This only manifests outside transaction contexts because
AbortTransaction() calls LWLockReleaseAll() during transaction
abort, releasing locks before DSM cleanup. Background workers and
other non-transactional code paths are vulnerable.
Fix by calling LWLockReleaseAll() unconditionally at the start of
shmem_exit(), before any callbacks run. Releasing locks before
callbacks prevents the segfault - locks must be released before
dsm_backend_shutdown() detaches their memory. This is safe because
after an error, held locks are protecting potentially inconsistent
data anyway, and callbacks can acquire fresh locks if needed.
Also add a comment noting that LWLockReleaseAll() must be safe to
call before LWLock initialization (which it is, since
num_held_lwlocks will be 0), plus an Assert for the post-condition.
This fix aligns with the original design intent from commit
001a573a2, which noted that backends must clean up shared memory
state (including releasing lwlocks) before unmapping dynamic shared
memory segments.
Reported-by: Rahila Syed <rahilasyed90@gmail.com>
Author: Rahila Syed <rahilasyed90@gmail.com>
Reviewed-by: Amit Langote <amitlangote09@gmail.com>
Reviewed-by: Dilip Kumar <dilipbalaut@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Discussion: https://postgr.es/m/CAH2L28uSvyiosL+kaic9249jRVoQiQF6JOnaCitKFq=xiFzX3g@mail.gmail.com
Backpatch-through: 14
On restart, a replica can fail with an error like 'unexpected data
beyond EOF in block 200 of relation T/D/R'. These are the steps to
reproduce it:
- A relation has a size of 400 blocks.
- Blocks 201 to 400 are empty.
- Block 200 has two rows.
- Blocks 100 to 199 are empty.
- A restartpoint is done
- Vacuum truncates the relation to 200 blocks
- A FPW deletes a row in block 200
- A checkpoint is done
- A FPW deletes the last row in block 200
- Vacuum truncates the relation to 100 blocks
- The replica restarts
When the replica restarts:
- The relation on disk starts at 100 blocks, because all the
truncations were applied before restart.
- The first truncate to 200 blocks is replayed. It silently fails, but
it will still (incorrectly!) update the cache size to 200 blocks
- The first FPW on block 200 is applied. XLogReadBufferForRead relies
on the cached size and incorrectly assumes that the page already
exists in the file, and thus won't extend the relation.
- The online checkpoint record is replayed, calling smgrdestroyall
which causes the cached size to be discarded
- The second FPW on block 200 is applied. This time, the detected size
is 100 blocks, an extend is attempted. However, the block 200 is
already present in the buffer cache due to the first FPW. This
triggers the 'unexpected data beyond EOF'.
To fix, update the cached size in SmgrRelation with the current size
rather than the requested new size, when the requested new size is
greater.
Author: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Discussion: https://www.postgresql.org/message-id/CAO6_Xqrv-snNJNhbj1KjQmWiWHX3nYGDgAc=vxaZP3qc4g1Siw@mail.gmail.com
Backpatch-through: 14
If a multixid with zero offset is left behind after a crash, and that
multixid later becomes the oldest multixid, truncation might try to
look up its offset and read the zero value. In the worst case, we
might incorrectly use the zero offset to truncate valid SLRU segments
that are still needed. I'm not sure if that can happen in practice, or
if there are some other lower-level safeguards or incidental reasons
that prevent the caller from passing an unwritten multixid as the
oldest multi. But better safe than sorry, so let's add an explicit
check for it.
In stable branches, we should perhaps do the same check for
'oldestOffset', i.e. the offset of the old oldest multixid (in master,
'oldestOffset' is gone). But if the old oldest multixid has an invalid
offset, the damage has been done already, and we would never advance
past that point. It's not clear what we should do in that case. The
check that this commit adds will prevent such an multixid with invalid
offset from becoming the oldest multixid in the first place, which
seems enough for now.
Reviewed-by: Andrey Borodin <x4mmm@yandex-team.ru>
Discussion: Discussion: https://www.postgresql.org/message-id/000301b2-5b81-4938-bdac-90f6eb660843@iki.fi
Backpatch-through: 14
The test 002_save_fullpage.pl, checking --save-fullpage fails with
wal_consistency_checking enabled, due to the fact that the block saved
in the file has the same LSN as the LSN used in the file name. The test
required that the block LSN is stritly lower than file LSN. This commit
relaxes the check a bit, by allowing the LSNs to match.
While on it, the test name is reworded to include some information about
the file and block LSNs, which is useful for debugging.
Author: Andrey Borodin <x4mmm@yandex-team.ru>
Discussion: https://postgr.es/m/4226AED7-E38F-419B-AAED-9BC853FB55DE@yandex-team.ru
Backpatch-through: 16
When creating a partition for a RANGE partitioned table, the reporting
of errors relating to converting the specified range values into
constant values for the partition key's type could display the name of a
previous partition key column when an earlier range was specified as
MINVALUE or MAXVALUE.
This was caused by the code not correctly incrementing the index that
tracks which partition key the foreach loop was working on after
processing MINVALUE/MAXVALUE ranges.
Fix by using foreach_current_index() to ensure the index variable is
always set to the List element being worked on.
Author: myzhen <zhenmingyang@yeah.net>
Reviewed-by: zhibin wang <killerwzb@gmail.com>
Discussion: https://postgr.es/m/273cab52.978.19b96fc75e7.Coremail.zhenmingyang@yeah.net
Backpatch-through: 14
A race condition could cause a newly created replication slot to become
invalidated between WAL reservation and a checkpoint.
Previously, if the required WAL was removed, we retried the reservation
process. However, the slot could still be invalidated before the retry if
the WAL was not yet removed but the checkpoint advanced the redo pointer
beyond the slot's intended restart LSN and computed the minimum LSN that
needs to be preserved for the slots.
The fix is to acquire an exclusive lock on ReplicationSlotAllocationLock
during WAL reservation, and a shared lock during the minimum LSN
calculation at checkpoints to serialize the process. This ensures that, if
WAL reservation occurs first, the checkpoint waits until restart_lsn is
updated before calculating the minimum LSN. If the checkpoint runs first,
subsequent WAL reservations pick a position at or after the latest
checkpoint's redo pointer.
We used a similar fix in HEAD (via commit 006dd4b2e5) and 18. The
difference is that in 17 and prior branches we need to additionally handle
the race condition with slot's minimum LSN computation during checkpoints.
Reported-by: suyu.cmj <mengjuan.cmj@alibaba-inc.com>
Author: Hou Zhijie <houzj.fnst@fujitsu.com>
Author: vignesh C <vignesh21@gmail.com>
Reviewed-by: Hayato Kuroda <kuroda.hayato@fujitsu.com>
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Backpatch-through: 14
Discussion: https://postgr.es/m/5e045179-236f-4f8f-84f1-0f2566ba784c.mengjuan.cmj@alibaba-inc.com
Commit c7eab0e97 changed the default password_encryption setting to
'scram-sha-256', so update the example for creating a user with an
assigned password.
In addition, commit 08951a7c9 added new options that in turn pass
default tokens NOBYPASSRLS and NOREPLICATION to the CREATE ROLE
command, so fix this omission as well for v16 and later.
Reported-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/cff1ea60-c67d-4320-9e33-094637c2c4fb%40iki.fi
Backpatch-through: 14
When processing the "publish" options of an ALTER PUBLICATION command,
we call SplitIdentifierString() to split the options into a List of
strings. Since SplitIdentifierString() modifies the delimiter
character and puts NULs in their place, this would overwrite the memory
of the AlterPublicationStmt. Later in AlterPublicationOptions(), the
modified AlterPublicationStmt is copied for event triggers, which would
result in the event trigger only seeing the first "publish" option
rather than all options that were specified in the command.
To fix this, make a copy of the string before passing to
SplitIdentifierString().
Here we also adjust a similar case in the pgoutput plugin. There's no
known issues caused by SplitIdentifierString() here, so this is being
done out of paranoia.
Thanks to Henson Choi for putting together an example case showing the
ALTER PUBLICATION issue.
Author: sunil s <sunilfeb26@gmail.com>
Reviewed-by: Henson Choi <assam258@gmail.com>
Reviewed-by: zengman <zengman@halodbtech.com>
Backpatch-through: 14
Prior to v15, GUC settings supplied in the CONNECTION clause of
CREATE SUBSCRIPTION were correctly passed through to
the publisher's walsender. For example:
CREATE SUBSCRIPTION mysub
CONNECTION 'options=''-c wal_sender_timeout=1000'''
PUBLICATION ...
would cause wal_sender_timeout to take effect on the publisher's walsender.
However, commit f3d4019da5d changed the way logical replication
connections are established, forcing the publisher's relevant
GUC settings (datestyle, intervalstyle, extra_float_digits) to
override those provided in the CONNECTION string. As a result,
from v15 through v18, GUC settings in the CONNECTION string were
always ignored.
This regression prevented per-connection tuning of logical replication.
For example, using a shorter timeout for walsender connecting
to a nearby subscriber and a longer one for walsender connecting
to a remote subscriber.
This commit restores the intended behavior by ensuring that
GUC settings in the CONNECTION string are again passed through
and applied by the walsender, allowing per-connection configuration.
Backpatch to v15, where the regression was introduced.
Author: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Chao Li <lic@highgo.com>
Reviewed-by: Kirill Reshke <reshkekirill@gmail.com>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Reviewed-by: Japin Li <japinli@hotmail.com>
Discussion: https://postgr.es/m/CAHGQGwGYV+-abbKwdrM2UHUe-JYOFWmsrs6=QicyJO-j+-Widw@mail.gmail.com
Backpatch-through: 15
This fixes a poorly written integer comparison function which was
performing subtraction in an attempt to return a negative value when
a < b and a positive value when a > b, and 0 when the values were equal.
Unfortunately that didn't always work correctly due to two's complement
having the INT_MIN 1 further from zero than INT_MAX. This could result
in an overflow and cause the comparison function to return an incorrect
result, which would result in the binary search failing to find the
value being searched for.
This could cause poor selectivity estimates when the statistics stored
the value of INT_MAX (2147483647) and the value being searched for was
large enough to result in the binary search doing a comparison with that
INT_MAX value.
Author: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CAEoWx2ng1Ot5LoKbVU-Dh---dFTUZWJRH8wv2chBu29fnNDMaQ@mail.gmail.com
Backpatch-through: 14
jit_profiling_support=true captures profile data for Linux perf. On
other platforms, LLVMCreatePerfJITEventListener() returns NULL and the
attempt to register the listener would crash.
Fix by ignoring the setting in that case. The documentation already
says that it only has an effect if perf support is present, and we
already did the same for older LLVM versions that lacked support.
No field reports, unsurprisingly for an obscure developer-oriented
setting. Noticed in passing while working on commit 1a28b4b4.
Backpatch-through: 14
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKGJgB6gvrdDohgwLfCwzVQm%3DVMtb9m0vzQn%3DCwWn-kwG9w%40mail.gmail.com
