Previously, ProcSignalInit() read the global barrier generation before
publishing its PID into pss_pid. This created a race condition: a
process could initialize its local generation with an older global
value, while a concurrent EmitProcSignalBarrier() might skip that
process because its pss_pid was still zero. This resulted in
WaitForProcSignalBarrier() hanging indefinitely.
Fix this by publishing pss_pid before reading psh_barrierGeneration
with a memory barrier so that the store to pss_pid is ordered before
the load. A concurrent EmitProcSignalBarrier() then either observes
the published PID and signals this slot, or completes its generation
increment before we load it.
While this race has become more visible due to recent features using
signal barriers in more places (such as online wal_level changes), the
issue is theoretically present since signal barriers were introduced
to release smgr caches (e.g., in DROP DATABASE). v14 has the
procsiangl barrier infrastricutre but no in-tree caller that actually
emits a barrier, so the case is unreachable there.
This issue was also reported by buildfarm member flaviventris.
Reported-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Alexander Lakhin <exclusion@gmail.com>
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://postgr.es/m/CAEze2WgAJmWReDN7Chtba8Er2YBvKCoa0KVN25-1evnTrHsLyA@mail.gmail.com
Backpatch-through: 15
Concurrent DDL can leave behind objects referencing other objects that
no longer exist. This can happen if an object is dropped, while a new
object that depends on it is created concurrently. For example:
session 1: BEGIN; CREATE FUNCTION myschema.myfunc() ...;
session 2: DROP SCHEMA myschema;
session 1: COMMIT;
DROP SCHEMA does check that there are no objects dependending on the
schema being dropped, but it does not see objects being concurrently
created by other sessions. Even if it did, this scenario would still
fail:
session 1: BEGIN: DROP SCHEMA myschema;
session 2: CREATE FUNCTION myschema.myfunc() ...;
session 1: COMMIT;
When the DROP SCHEMA runs, the schema was empty, but the new function
is created in it before the dropping transaction completes. The CREATE
FUNCTION does not see that the schema is concurrently being dropped.
In both of these scenarios, the function is left behind in the schema
that no longer exists.
To fix, acquire AccessShareLock on all referenced objects when
recording dependencies. This conflicts with the AccessExclusiveLock
taken by DROP, preventing the race. After acquiring the lock, verify
that the object still exists, and if it was dropped concurrently,
report an error. We already had such a mechanism for shared
dependencies, but for some reason we didn't do it for in-database
dependendies.
Ideally the locks would be acquired much earlier when creating a new
object, but that will require modifying a lot of callers. This check
while recording the dependency is a nice wholesale protection, and
even if we change all the CREATE commands to acquire locks earlier,
it's still good to have this as a backstop to catch any cases where we
forgot to do so.
The patch adds a few tests for some cases that left behind orphaned
objects before this. It also adds a test for roles, which already had
such protection, although that test is partially disabled because the
error message includes an OID which is not predictable.
Author: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Reviewed-by: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Discussion: https://postgr.es/m/ZiYjn0eVc7pxVY45@ip-10-97-1-34.eu-west-3.compute.internal
Backpatch-through: 14
When creating a relation with a dropped column, we called
recordDependencyOn() also on the datatype of the dropped column, which
is always InvalidOid. In versions 15 and above, that was harmless
because recordDependencyOn() considers InvalidOid as a pinned object,
and skips over it. On version 14, isPinnedObject() does not consider
InvalidOid as pinned, so we created a bogus pg_depend entry with
refobjectid == 0.
As far as I can tell, the only case when AddNewAttributeTuples() is
called with dropped columns is when performing a table-rewriting ALTER
TABLE command. That temporarily creates a new relation with the same
columns, including dropped ones, then swaps the relations, and drops
the newly created table again. So even on version 14, the bogus
pg_depend entry was only on the transient relation that was dropped at
the end of the ALTER TABLE command, which was harmless.
Even though this is harmless, let's be tidy, similar to commit
713bce9484. The reason I noticed this now and why I backported this,
is because the next commit will add code to acquire locks on the
referenced objects, and we don't want to acquire a lock on InvalidOid.
Discussion: https://postgr.es/m/ZiYjn0eVc7pxVY45@ip-10-97-1-34.eu-west-3.compute.internal
Backpatch-through: 14
Commit 77dff5d937 introduced a SimpleLruWriteAll() call when replaying
multixact WAL records generated by older minor versions. However,
SimpleLruWriteAll() acquires the SLRU lock and on v16 and below, it's
called while already holding the lock, leading to self-deadlock.
Version 17 and 18 did not have that problem, because in those versions
the lock is acquired later in the function.
To fix, acquire MultiXactOffsetSLRULock later in RecordNewMultiXact(),
at the same place where it's acquired on version 17 and 18.
Author: Andrey Borodin <x4mmm@yandex-team.ru>
Reported-by: Radim Marek <radim@boringsql.com>
Discussion: https://www.postgresql.org/message-id/19490-9c59c6a583513b99@postgresql.org
Backpatch-through: 14-16
DisownLatch() was executed after the PGPROC entry of the process
terminated is pushed back into a freelist. A newly-forked backend that
recycles the slot could call OwnLatch() and PANIC with a "latch already
owned by PID", taking down the server.
There were two scenarios related to lock groups where this issue could
be reached:
* A follower pushes the leader's PGPROC back to the freelist while the
leader has not yet called DisownLatch() in its own ProcKill().
* A leader outliving all its followers pushes its own PGPROC onto the
freelist before reaching DisownLatch(), which would be the most common
scenario.
This issue is fixed by calling SwitchBackToLocalLatch() and
DisownLatch() at an earlier phase of ProcKill(), before any freelist
manipulation happens, so that the slot of the backend terminated is
never exposed as owning a latch.
Note that pgstat_reset_wait_event_storage() is kept at a later stage.
An upcoming commit will take advantage of that by introducing a test
able to check the original PANIC scenario.
Author: Vlad Lesin <vladlesin@gmail.com>
Reviewed-by: Andrey Borodin <x4mmm@yandex-team.ru>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/d2983796-2603-41b7-a66e-fc8489ddb954@gmail.com
Backpatch-through: 14
This commit fixes two bugs in ProcKill()'s lock-group teardown freelist
publication:
* a double push of the leader's PGPROC that corrupts the freelist.
* a leak of the last follower's PGPROC slot.
ProcKill()'s lock-group teardown had two PGPROC freelist updates
scattered through the function, done under two separate freeProcsLock
acquisitions:
* A follower's push of the leader's PGPROC, done when a follower is the
last group member exiting.
* Every backend's self-push at the bottom of the function.
The two freelist updates were coordinated only by inspecting
proc->lockGroupLeader, which a follower could clear as a side effect of
pushing the leader. This coordination was broken. For example, with
two concurrent backends:
* The follower clears leader->lockGroupLeader and pushes the leader's
PGPROC under leader_lwlock.
* The follower does not clear its own proc->lockGroupLeader, being
skipped.
* When the leader reaches the bottom of ProcKill(), it sees a NULL
proc->lockGroupLeader (the follower cleared it) and pushes itself,
causing a second dlist_push_tail() of the same node onto the same
freelist.
* The follower at the bottom sees its own proc->lockGroupLeader being
not NULL (never cleared) and skips its own push, causing its own slot
to leak.
This commit refactors the freelist manipulation to be done in two
distinct phases, each step using its own lock acquisition to ensure that
each freelist operation happens in an isolated manner for each backend
(follower or leader):
- First, under a single leader_lwlock acquisition, check the state of
the lock-group. Depending on if we are dealing with a follower and/or a
leader, and if the leader has exited before a follower, then set some
state booleans that define which actions should be taken with the
freelist.
- Second, under a single freeProcsLock acquisition, perform the cleanup
actions, self-push of a backend and/or push of the leader back to the
freelist.
This is an old issue, dating back to 9.6 where parallel workers and lock
grouping has been added.
Author: Vlad Lesin <vladlesin@gmail.com>
Reviewed-by: Andrey Borodin <x4mmm@yandex-team.ru>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/d2983796-2603-41b7-a66e-fc8489ddb954@gmail.com
Backpatch-through: 14
Given a WHERE clause like "int[] @@ query_int" or "query_int ~~ int[]"
where the query_int side is a table column having statistics,
_int_matchsel() exited without remembering to free the statistics
tuple. This would typically lead to warnings about cache refcount
leakage, like
WARNING: resource was not closed: cache pg_statistic (73), tuple 42/12 has count 1
It's been wrong since this code was added, in commit c6fbe6d6f.
Bug: #19492
Reported-by: Man Zeng <zengman@halodbtech.com>
Author: Man Zeng <zengman@halodbtech.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/19492-ddcd0e22399ef85a@postgresql.org
Backpatch-through: 14
The check for the minimum expected bytea size of a MVDependencies object
was using SizeOfItem() for its calculation. This macro uses the number
of attributes in a single dependency.
This minimum size calculation should be based on MinSizeOfItems(), that
computes the minimum expected size as the header plus the
minimally-sized number of dependency items.
Oversight in d08c44f7a4.
Author: Ilia Evdokimov <ilya.evdokimov@tantorlabs.com>
Discussion: https://postgr.es/m/4b8d299d-2505-4c30-bf80-0f697410db35@tantorlabs.com
Backpatch-through: 14
When reusing an existing WAL receiver after it has reached
WALRCV_WAITING for new instructions, RequestXLogStreaming() copied
PrimaryConnInfo into WalRcv->conninfo before switching the state to
WALRCV_RESTARTING. At that point ready_to_display could still be true,
so pg_stat_wal_receiver could expose the raw connection string,
including sensitive fields, but it should only show the user-displayable
version of the connection string.
WALRCV_RESTARTING does not establish a new connection. The waiting WAL
receiver reuses its existing connection and only needs a new startpoint
and timeline, so there is no need to copy the raw connection string into
shared memory again. Let's only copy conninfo when launching a new WAL
receiver after WALRCV_STOPPED, not while waiting for instructions.
This commit adds coverage for the case fixed by this commit to the
timeline-switch test by verifying that the WAL receiver conninfo remains
consistent across the jump.
Backpatch all the way down, as this issue is possible since
pg_stat_wal_receiver has been introduced.
Author: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/EF91FF76-1E2B-4F3B-9162-290B4DC517FF@gmail.com
Backpatch-through: 14
Commit c37b3d08ca attempted to preserve group permissions on pg_recvlogical
output files when group access was enabled on the source cluster. However,
the output files were still created with a fixed S_IRUSR | S_IWUSR mode,
preventing group-read permissions from being applied.
This commit fixes the issue by creating output files with pg_file_create_mode
instead of a hard-coded mode. This allows pg_recvlogical to correctly preserve
group permissions from the source cluster.
Backpatch to all supported branches.
Author: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Srinath Reddy Sadipiralla <srinath2133@gmail.com>
Discussion: https://postgr.es/m/CAHGQGwHhpizYzMo3nFP4GkNMueSNMY3QfC-gBN1VTXtuiANDvw@mail.gmail.com
Backpatch-through: 14
Three locations use Assert() to guard against a mismatch between the
number of columns advertised in the RELATION message and the number
actually received in the subsequent INSERT/UPDATE tuple message. Since
these values originate from the publisher, the check must survive into
production builds.
A malicious or buggy publisher can send a RELATION claiming N columns
and an INSERT claiming M < N columns. The subscriber's apply worker
indexes into colvalues[]/colstatus[] using column indices from the
RELATION message's attribute map, causing a heap out-of-bounds read when
the tuple's column array is smaller than expected. We've looked, without
success, for a scenario in which the publisher holds sufficient control
over these out-of-bounds bytes to exploit this or even to reach a
SIGSEGV. Despite not finding one, the code has been fragile. Back-patch
to v14 (all supported versions).
Reported-by: Varik Matevosyan <varikmatevosyan@gmail.com>
Author: Varik Matevosyan <varikmatevosyan@gmail.com>
Discussion: https://postgr.es/m/CA+bBoog3cCogktzfLb9bppUByu-10B3CFp8u=iKXG_OvtAguCw@mail.gmail.com
Backpatch-through: 14
These tests have been removed by 906ea101d0, due to some of them being
unstable in the buildfarm with low max_stack_depth values. They are now
reworked so as they should be more portable.
The tests to cover the findoprnd() overflows use a balanced tree to
avoid using too much stack, per a suggestion and an investigation by Tom
Lane.
Note: This is initially applied only on HEAD; a backpatch will follow
should the buildfarm be fine with the situation.
Discussion: https://postgr.es/m/agZc6XecyE7E7fep@paquier.xyz
Backpatch-through: 14
When an UPDATE statement triggers check_foreign_key() with the
action set to "cascade", it generates more UPDATE statements to
modify the key values in referencing relations. If a new key value
is NULL, SPI_getvalue() returns a NULL pointer, which is
subsequently passed to quote_literal_cstr(), causing a segfault.
To fix, skip quoting when a new key value is NULL and insert an
unquoted NULL keyword instead.
Oversight in commit 260e97733b. While the refint documentation
recommends marking primary key columns NOT NULL, the aforementioned
scenario accidentally worked on platforms where snprintf()
substitutes "(null)" for NULL pointers. Note that for
character-type columns, the old code quoted "(null)" as a string
literal, so this didn't always produce correct results. But it
still seems better to fix this than to reject cases that previously
worked.
Reported-by: Nikita Kalinin <n.kalinin@postgrespro.ru>
Author: Ayush Tiwari <ayushtiwari.slg01@gmail.com>
Reviewed-by: Pierre Forstmann <pierre.forstmann@gmail.com>
Discussion: https://postgr.es/m/19476-bd04ea6241345303%40postgresql.org
Backpatch-through: 14
When pgbench runs with multiple threads and verbose error reporting is
enabled (--verbose-errors), multiple clients can build verbose error
messages concurrently. Previously, a function-local static
PQExpBuffer was used for these messages, causing the buffer to be
shared across threads. This was not thread-safe and could result in
corrupted or incorrect log output.
Fix this by using a local PQExpBufferData instead of a static buffer.
This keeps verbose error messages correct during concurrent execution.
Backpatch to v15, where this issue was introduced.
Author: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Alex Guo <guo.alex.hengchen@gmail.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Discussion: https://postgr.es/m/CAHGQGwER1AjGXpkKB9t9820NBhMQ_Ghv7=HsKeodUr3=SZsF4g@mail.gmail.com
Backpatch-through: 15
Two cases fixed by 2b5ba2a0a1 were not covered, to emulate the
handling of corrupted data, for:
- set control bit with a valid 2-byte match tag where offset is 0.
- set control bit with a valid 2-byte match tag where offset exceeds
output written.
Oversight in 67d318e704.
Reviewed-by: Ayush Tiwari <ayushtiwari.slg01@gmail.com>
Discussion: https://postgr.es/m/agF4xkIdRcrCIprs@paquier.xyz
Backpatch-through: 14
Previously, pg_stat_progress_copy in the subscriber could continue to show
the initial COPY operation for logical replication table synchronization as
active even after the data copy had finished. The stale progress entry
remained visible until synchronization caught up with the publisher.
This happened because the table synchronization code called BeginCopyFrom()
and CopyFrom(), but failed to call EndCopyFrom() afterward.
This commit fixes the issue by adding the missing EndCopyFrom() call so that
the COPY progress state in the subscriber is cleared as soon as the initial
data copy completes.
Backpatch to all supported branches.
Author: Shinya Kato <shinya11.kato@gmail.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: ChangAo Chen <cca5507@qq.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Discussion: https://postgr.es/m/CAOzEurQKuy3RiPkd=25PEwEzaqHuGvEOf=X7vaVzhgNjaukYzA@mail.gmail.com
Backpatch-through: 14
The postmaster test 004_negotiate.pl could fail due to IO::Socket::INET
gone missing, in environments that cannot use Unix sockets.
Oversight in the backport done in 6dffaeb8e5, so like the other commit
this is applied across the v14~17 range. Per buildfarm member drongo.
Security: CVE-2026-6479
Backpatch-through: 14
These checks are failing in the buildfarm, reporting stack overflows
rather than the expected errors, though seemingly only on ppc64 and
s390x platforms. Perhaps there is something off about our tests
for stack depth on those architectures? But there's no time to
debug that right now, and surely these tests aren't too essential.
Revert for now and plan to revisit after the release dust settles.
Backpatch-through: 14
Security: CVE-2026-6473
Maliciously crafted key value updates could achieve SQL injection
within check_foreign_key(). To fix, ensure new key values are
properly quoted and escaped in the internally generated SQL
statements. While at it, avoid potential buffer overruns by
replacing the stack buffers for internally generated SQL statements
with StringInfo.
Reported-by: Nikolay Samokhvalov <nik@postgres.ai>
Author: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Security: CVE-2026-6637
Backpatch-through: 14
When result_is_int is set to 0, PQfn() cannot validate that the
result fits in result_buf, so it will write data beyond the end of
the buffer when the server returns more data than requested. Since
this function is insecurable and obsolete, add a warning to the top
of the pertinent documentation advising against its use.
The only in-tree caller of PQfn() is the frontend large object
interface. To fix that, add a buf_size parameter to
pqFunctionCall3() that is used to protect against overruns, and use
it in a private version of PQfn() that also accepts a buf_size
parameter.
Reported-by: Yu Kunpeng <yu443940816@live.com>
Reported-by: Martin Heistermann <martin.heistermann@unibe.ch>
Author: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Etsuro Fujita <etsuro.fujita@gmail.com>
Security: CVE-2026-6477
Backpatch-through: 14
If you accumulate many arrays full of NULLs, you could overflow
'nitems', before reaching the MaxAllocSize limit on the allocations.
Add an explicit check that the number of items doesn't grow too large.
With more than MaxArraySize items, getting the final result with
makeArrayResultArr() would fail anyway, so better to error out early.
Reported-by: Xint Code
Author: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Backpatch-through: 14
Security: CVE-2026-6473
pg_locale_icu.c was full of places where a very long input string
could cause integer overflow while calculating a buffer size,
leading to buffer overruns.
It also was cavalier about using char-type local arrays as buffers
holding arrays of UChar. The alignment of a char[] variable isn't
guaranteed, so that this risked failure on alignment-picky platforms.
The lack of complaints suggests that such platforms are very rare
nowadays; but it's likely that we are paying a performance price on
rather more platforms. Declare those arrays as UChar[] instead,
keeping their physical size the same.
pg_locale_libc.c's strncoll_libc_win32_utf8() also had the
disease of assuming it could double or quadruple the input
string length without concern for overflow.
Reported-by: Xint Code
Reported-by: Pavel Kohout <pavel.kohout@aisle.com>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Backpatch-through: 14
Security: CVE-2026-6473
pg_rewind and pg_basebackup could be fed paths from rogue endpoints that
could overwrite the contents of the client when received, achieving path
traversal.
There were two areas in the tree that were sensitive to this problem:
- pg_basebackup, through the astreamer code, where no validation was
performed before building an output path when streaming tar data. This
is an issue in v15 and newer versions.
- pg_rewind file operations for paths received through libpq, for all
the stable branches supported.
In order to address this problem, this commit adds a helper function in
path.c, that reuses path_is_relative_and_below_cwd() after applying
canonicalize_path(). This can be used to validate the paths received
from a connection point. A path is considered invalid if any of the two
following conditions is satisfied:
- The path is absolute.
- The path includes a direct parent-directory reference.
Reported-by: XlabAI Team of Tencent Xuanwu Lab
Reported-by: Valery Gubanov <valerygubanov95@gmail.com>
Author: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Backpatch-through: 14
Security: CVE-2026-6475
A few functions in this file were incautious about multiplying a
possibly large integer by a factor more than 1 and then using it as
an allocation size. This is harmless on 64-bit systems where we'd
compute a size exceeding MaxAllocSize and then fail, but on 32-bit
systems we could overflow size_t, leading to an undersized
allocation and buffer overrun. To fix, use palloc_array() or
mul_size() instead of handwritten multiplication.
Reported-by: Sven Klemm <sven@tigerdata.com>
Reported-by: Xint Code
Author: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Tatsuo Ishii <ishii@postgresql.org>
Security: CVE-2026-6473
Backpatch-through: 14
This omission allowed roles to create multirange types in any
schema, potentially leading to privilege escalations. Note that
when a multirange type name is not specified in CREATE TYPE, it is
automatically placed in the range type's schema, which is checked
at the beginning of DefineRange().
Reported-by: Jelte Fennema-Nio <postgres@jeltef.nl>
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Security: CVE-2026-6472
Backpatch-through: 14
Although pg_strftime() has defined error conditions, no callers bother
to check for errors. This is problematic because the output string is
very likely not null-terminated if an error occurs, so that blindly
using it is unsafe. Rather than trusting that we can find and fix all
the callers, let's alter the function's API spec slightly: make it
guarantee a null-terminated result so long as maxsize > 0.
Furthermore, if we do get an error, let's make that null-terminated
result be an empty string. We could instead truncate at the buffer
length, but that risks producing mis-encoded output if the tz_name
string contains multibyte characters. It doesn't seem reasonable for
src/timezone/ to make use of our encoding-aware truncation logic.
Also, the only really likely source of a failure is a user-supplied
timezone name that is intentionally trying to overrun our buffers.
I don't feel a need to be particularly friendly about that case.
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: John Naylor <johncnaylorls@gmail.com>
Backpatch-through: 14
Security: CVE-2026-6474
timeofday() assumed that the output of pg_strftime() could not contain
% signs, other than the one it explicitly asks for with %%. However,
we don't have that guarantee with respect to the time zone name (%Z).
A crafted time zone setting could abuse the subsequent snprintf()
call, resulting in crashes or disclosure of server memory.
To fix, split the pg_strftime() call into two and then treat the
outputs as literal strings, not a snprintf format string. The
extra pg_strftime() call doesn't really cost anything, since the
bulk of the conversion work was done by pg_localtime().
Also, adjust buffer widths so that we're not risking string truncation
during the snprintf() step, as that would create a hazard of producing
mis-encoded output.
This also fixes a latent portability issue: the format string expects
an int, but tp.tv_usec is long int on many platforms.
Reported-by: Xint Code
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: John Naylor <johncnaylorls@gmail.com>
Backpatch-through: 14
Security: CVE-2026-6474
This commit applies timingsafe_bcmp() to authentication paths that
handle attributes or data previously compared with memcpy() or strcmp(),
which are sensitive to timing attacks.
The following data is concerned by this change, some being in the
backend and some in the frontend:
- For a SCRAM or MD5 password, the computed key or the MD5 hash compared
with a password during a plain authentication.
- For a SCRAM exchange, the stored key, the client's final nonce and the
server nonce.
- RADIUS (up to v18), the encrypted password.
- For MD5 authentication, the MD5(MD5()) hash.
Reported-by: Joe Conway <mail@joeconway.com>
Security: CVE-2026-6478
Author: Michael Paquier <michael@paquier.xyz>
Reviewed-by: John Naylor <johncnaylorls@gmail.com>
Backpatch-through: 14
timingsafe_bcmp() should be used instead of memcmp() or a naive
for-loop, when comparing passwords or secret tokens, to avoid leaking
information about the secret token by timing. This commit just
introduces the function but does not change any existing code to use
it yet.
This has been initially applied as of 09be391126 in v18 and newer
versions, and will be used in all the stable branches for an upcoming
fix.
Co-authored-by: Jelte Fennema-Nio <github-tech@jeltef.nl>
Discussion: https://www.postgresql.org/message-id/7b86da3b-9356-4e50-aa1b-56570825e234@iki.fi
Security: CVE-2026-6478
Backpatch-through: 14
contrib/intarray's query_int type uses an int16 field to hold the
offset from a binary operator node to its left operand. However, it
allows the number of nodes to be as much as will fit in MaxAllocSize,
so there is a risk of overflowing int16 depending on the precise shape
of the tree. Simple right-associative cases like "a | b | c | ..."
work fine, so we should not solve this by restricting the overall
number of nodes. Instead add a direct test of whether each individual
offset is too large.
contrib/ltree's ltxtquery type uses essentially the same logic and
has the same 16-bit restriction.
(The core backend's tsquery.c has a variant of this logic too, but
in that case the target field is 32 bits, so it is okay so long
as varlena datums are restricted to 1GB.)
In v16 and up, these types support soft error reporting, so we have
to complicate the recursive findoprnd function's API a bit to allow
the complaint to be reported softly. v14/v15 don't need that.
Undocumented and overcomplicated code like this makes my head hurt,
so add some comments and simplify while at it.
Reported-by: Xint Code
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Backpatch-through: 14
Security: CVE-2026-6473
Define MaxAllocSize in src/include/common/fe_memutils.h rather
than having several copies of it in different src/common/*.c files.
This also provides an opportunity to document it better.
Back-patch of commit 11b7de4a7, needed now because assorted security
fixes are adding additional references to MaxAllocSize in frontend
code.
Backpatch-through: 14-17
Security: CVE-2026-6473
The handling of SSL and GSS negotiation messages in
ProcessStartupPacket() could cause a recursion of the backend,
ultimately crashing the server as the negotiation attempts were not
tracked across multiple calls processing startup packets.
A malicious client could therefore alternate rejected SSL and GSS
requests indefinitely, each adding a stack frame, until the backend
crashed with a stack overflow, taking down a server.
This commit addresses this issue by modifying ProcessStartupPacket() so
as processed negotiation attempts are tracked, preventing infinite
recursive attempts. A TAP test is added to check this problem, where
multiple SSL and GSS negotiated attempts are stacked.
Reported-by: Calif.io in collaboration with Claude and Anthropic
Research
Author: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Security: CVE-2026-6479
Backpatch-through: 14
These two routines will be used in a test of an upcoming fix. This
commit affects the v14~v17 range. v18 and newer versions already
include them, thanks to 85ec945b78.
Security: CVE-2026-6479
Backpatch-through: 14
multirange_recv and BlockRefTableReaderNextRelation were incautious
about multiplying a possibly-large integer by a factor more than 1
and then using it as an allocation size. This is harmless on 64-bit
systems where we'd compute a size exceeding MaxAllocSize and then
fail, but on 32-bit systems we could overflow size_t leading to an
undersized allocation and buffer overrun.
Fix these places by using palloc_array() instead of a handwritten
multiplication. (In HEAD, some of them were fixed already, but
none of that work got back-patched at the time.)
In addition, BlockRefTableReaderNextRelation passes the same value
to BlockRefTableRead's "int length" parameter. If built for
64-bit frontend code, palloc_array() allows a larger array size
than it otherwise would, potentially allowing that parameter to
overflow. Add an explicit check to forestall that and keep the
behavior the same cross-platform.
Reported-by: Xint Code
Author: Tom Lane <tgl@sss.pgh.pa.us>
Backpatch-through: 14
Security: CVE-2026-6473
Some UTF8 characters decompose to more than a dozen codepoints.
It is possible for an input string that fits into well under
1GB to produce more than 4G decomposed codepoints, causing
unicode_normalize()'s decomp_size variable to wrap around to a
small positive value. This results in a small output buffer
allocation and subsequent buffer overrun.
To fix, test after each addition to see if we've overrun MaxAllocSize,
and break out of the loop early if so. In frontend code we want to
just return NULL for this failure (treating it like OOM). In the
backend, we can rely on the following palloc() call to throw error.
I also tightened things up in the calling functions in varlena.c,
using size_t rather than int and allocating the input workspace
with palloc_array(). These changes are probably unnecessary
given the knowledge that the original input and the normalized
output_chars array must fit into 1GB, but it's a lot easier to
believe the code is safe with these changes.
Reported-by: Xint Code
Reported-by: Bruce Dang <bruce@calif.io>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Co-authored-by: Heikki Linnakangas <hlinnaka@iki.fi>
Backpatch-through: 14
Security: CVE-2026-6473
The number of NFA states, number of NFA arcs, and number of colors
are all bounded to reasonably small values. However, there are
places where we try to allocate arrays sized by products of those
quantities, and those calculations could overflow, enabling
buffer-overrun attacks. In practice there's no problem on 64-bit
machines, but there are some live scenarios on 32-bit machines.
A related problem is that citerdissect() and creviterdissect()
allocate arrays based on the length of the input string, which
potentially could overflow.
To fix, invent MALLOC_ARRAY and REALLOC_ARRAY macros that rely on
palloc_array_extended and repalloc_array_extended with the NO_OOM
option, similarly to the existing MALLOC and REALLOC macros.
(Like those, they'll throw an error not return a NULL result for
oversize requests. This doesn't really fit into the regex code's
view of error handling, but it'll do for now. We can consider
whether to change that behavior in a non-security follow-up patch.)
I installed similar defenses in the colormap construction code.
It's not entirely clear whether integer overflow is possible
there, but analyzing the behavior in detail seems not worth
the trouble, as the risky spots are not in hot code paths.
I left a bunch of calls as-is after verifying that they can't
overflow given reasonable limits on nstates and narcs. Those
limits were enforced already via REG_MAX_COMPILE_SPACE, but
add commentary to document the interactions.
In passing, also fix a related edge case, which is that the
special color numbers used in LACON carcs could overflow the
"color" data type, if ncolors is close to MAX_COLOR.
In v14 and v15, the regex engine calls malloc() directly instead
of using palloc(), so MALLOC_ARRAY and REALLOC_ARRAY do likewise.
Reported-by: Xint Code
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
Backpatch-through: 14
Security: CVE-2026-6473
Sufficiently large "count" arguments could result in undetected
overflow, causing the allocated memory chunk to be much smaller
than what the caller will subsequently write into it. This is
unlikely to be a hazard with 64-bit size_t but can sometimes
happen on 32-bit builds, primarily where a function allocates
workspace that's significantly larger than its input data.
Rather than trying to patch the at-risk callers piecemeal,
let's just redefine these macros so that they always check.
To do that, move the longstanding add_size() and mul_size() functions
into palloc.h and mcxt.c, and adjust them to not be specific to
shared-memory allocation. Then invent palloc_mul(), palloc0_mul(),
palloc_mul_extended() to use these functions. Actually, the latter
use inlined copies to save one function call. repalloc_array() gets
similar treatment. I didn't bother trying to inline the calls for
repalloc0_array() though.
In v14 and v15, this also adds repalloc_extended(), which previously
was only available in v16 and up.
We need copies of all this in fe_memutils.[hc] as well, since that
module also provides palloc_array() etc.
Reported-by: Xint Code
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
Backpatch-through: 14
Security: CVE-2026-6473
Commit 600086f47 added (several bespoke copies of) size_t addition with
overflow checks to libpq. Move this to common/int.h, along with
its subtraction and multiplication counterparts.
pg_neg_size_overflow() is intentionally omitted; I'm not sure we should
add SSIZE_MAX to win32_port.h for the sake of a function with no
callers.
Back-patch of commit 8934f2136, done now because pg_add_size_overflow()
and friends are needed more widely for security fixes.
Author: Jacob Champion <jacob.champion@enterprisedb.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAOYmi%2B%3D%2BpqUd2MUitvgW1pAJuXgG_TKCVc3_Ek7pe8z9nkf%2BAg%40mail.gmail.com
Backpatch-through: 14-18
Security: CVE-2026-6473
The options "StartSel", "StopSel" and "FragmentDelimiter" given by a
caller of the SQL function ts_headline() have their lengths stored as
int16. When providing values larger than PG_INT16_MAX, it was possible
to overflow the length values stored, leading to incorrect behaviors in
generateHeadline(), in most cases translating to a crash.
Attempting to use values for these options larger than PG_INT16_MAX is
now blocked. Some test cases are added to cover our tracks.
Reported-by: Xint Code
Author: Michael Paquier <michael@paquier.xyz>
Backpatch-through: 14
Security: CVE-2026-6473
The lquery parser in contrib/ltree/ had two overflow problems:
- A single lquery level with many OR-separated variants (e.g.,
'label1|label2|...'), could cause an overflow of totallen, this being
stored as a uint16, meaning a maximum value of UINT16_MAX or 65k. Each
variant contributes MAXALIGN(LVAR_HDRSIZE + len) bytes. With enough
long variants, the value would wraparound. This would corrupt the data
written by LQL_NEXT(), leading to a stack corruption, most likely
translating into a crash, but it would allow incorrect memory access.
- numvar, labelled as a uint16, counts the number of OR-variants in a
single level, and it is incremented without bounds checking. With more
than PG_UINT16_MAX (65k) variants in a single level, and a minimum of
131kB of input data, it would wrap to 0. When a (wildcard) '*' is
used, this would change the query results silently.
For both issues, a set of overflows checks are added to guard against
these problematic patterns.
The first issue has been reported by the three people listed below,
affecting v16 and newer versions due to b1665bf01e. Its coding was
still unsafe in v14 and v15. The second issue affects all the stable
branches; I have bumped into while reviewing the code of the module.
Reported-by: Vergissmeinnicht <vergissmeinnichtzh@gmail.com>
Reported-by: A1ex <alex000young@gmail.com>
Reported-by: Jihe Wang <wangjihe.mail@gmail.com>
Author: Michael Paquier <michael@paquier.xyz>
Security: CVE-2026-6473
Backpatch-through: 14
As connections that failed abort cleanup can't safely be further used,
if a remote query tries to get such a connection, we reject it.
Previously, this rejection involved dropping the connection if it was
open, without accounting for the possibility of open cursors using it,
causing a server crash when such an open cursor tried to use an
already-dropped connection, as a cursor-handling function
(create_cursor, fetch_more_data, or close_cursor) was called on a freed
PGconn. To fix, delay dropping failed connections until abort cleanup
of the main transaction, to ensure open cursors using such a connection
can safely refer to the PGconn for it.
Oversight in commit 8bf58c0d9.
Reported-by: Zhibai Song <songzhibai1234@gmail.com>
Diagnosed-by: Zhibai Song <songzhibai1234@gmail.com>
Author: Etsuro Fujita <etsuro.fujita@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Discussion: https://postgr.es/m/CAPmGK176y6JP017-Cn%2BhS9CEJx_6iVhRoYbAqzuLU4d8-XPPNg%40mail.gmail.com
Backpatch-through: 14
rel_is_distinct_for()'s RTE_SUBQUERY branch passed only the equality
operator from each join clause to query_is_distinct_for(), discarding
the operator's input collation. query_is_distinct_for() then verified
opfamily compatibility but never checked collations, so a DISTINCT /
GROUP BY / set-op operating under one collation was trusted to prove
uniqueness for a comparison performed under an unrelated collation.
As with the recent fix in relation_has_unique_index_for(), this is
unsound for nondeterministic collations and yields wrong query results
in any optimization that consumes the proof.
Fix by carrying each clause's operator input collation into
query_is_distinct_for() and validating it at every check-site against
the subquery target expression's collation.
Back-patch to all supported branches. query_is_distinct_for() is
declared in an installed header, so on stable branches the existing
two-list signature is retained as a thin wrapper that forwards to a
new collation-aware entry point; external callers continue to receive
the historical collation-blind answer.
Author: Richard Guo <guofenglinux@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CAMbWs4_XUUSTyzCaRjUeeahWNqi=8ZOA5Q4coi8zUVEDSBkM6A@mail.gmail.com
Backpatch-through: 14
relation_has_unique_index_for() has long had an XXX noting that it
doesn't check collations when matching a unique index's columns
against equality clauses. This was benign as long as all collations
in play reduced to the same notion of equality, but has been incorrect
since nondeterministic collations were introduced in PG 12: a unique
index under a deterministic collation does not prove uniqueness under
a nondeterministic collation, nor vice versa.
The consequence is wrong query results for any planner optimization
that consumes the faulty proof, including inner-unique join execution
(which stops the inner search after the first match per outer row),
useless-left-join removal, semijoin-to-innerjoin reduction, and
self-join elimination.
Fix by requiring the index's collation to agree on equality with the
clause's input collation. Two collations agree on equality if either
is InvalidOid (denoting a non-collation-sensitive operation, which
cannot conflict with the other side), if they have the same OID, or if
both are deterministic: by definition a deterministic collation treats
two strings as equal iff they are byte-wise equal (see CREATE
COLLATION), so any two deterministic collations share the same
equality relation and the uniqueness proof carries over. Any mismatch
involving a nondeterministic collation is rejected.
Back-patch to all supported branches; the bug has existed since
nondeterministic collations were introduced in PG 12.
Author: Richard Guo <guofenglinux@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CAMbWs4_XUUSTyzCaRjUeeahWNqi=8ZOA5Q4coi8zUVEDSBkM6A@mail.gmail.com
Backpatch-through: 14
The XLogRecordPageWithFreeSpace function updates the freespace map (FSM) data
while replaying data-level WAL records during the recovery. If the FSM block
is updated, it needs to be marked as modified. Currently, this is done with
the MarkBufferDirtyHint call (as in all other cases for modifying FSM data).
However, in the recovery context, this function will actually do nothing if
checksums are enabled. It's assumed that the page should not be dirtied
during recovery while modifying hints to protect against torn pages, since no
new WAL data can be generated at this point to store FPI.
Such logic does not seem fully aligned with the FSM case, as its blocks could
be simply zeroed if a checksum mismatch is detected. Currently, changes to an
FSM block could be lost if each change to that block occurs infrequently
enough to allow it to be evicted from the cache. To persist the change, the
modification needs to be performed while the FSM block is still kept in
buffers and marked as dirty after receiving its FPI. If the block has already
been cleaned, the change won't be persisted, so stored FSM blocks may remain
in an obsolete state.
If a large number of discrepancies between the data in leaf FSM blocks and the
actual data blocks accumulate on the replica server, this could cause
significant delays in insert operations after switchover. Such an insert
operation may need to visit many data blocks marked as having sufficient
space in the FSM, only to discover that the information is incorrect and the
FSM records need to be corrected. In a heavily trafficked insert-only table
with many concurrent clients performing inserts, this has been observed to
cause several-second stalls, causing visible application malfunction. The
desire to avoid such cases was the reason behind the commit ab7dbd681, which
introduced an update of FSM data during the heap_xlog_visible invocation.
However, an update to the FSM data on the standby side could be lost due to a
missing 'dirty' flag, so there is still a possibility that a large number of
FSM records will contain incorrect data. Note that having a zeroed FSM page
in such a case (due to a checksum mismatch) is preferable, as a zero value
will be interpreted as an indication of full data blocks, and the inserter
will be routed to the next FSM block or to the end of the table.
Given that FSM is ready to handle torn page writes and
XLogRecordPageWithFreeSpace is called only during the recovery, there seems
to be no reason to use MarkBufferDirtyHint here instead of a regular
MarkBufferDirty call.
Discussion: https://postgr.es/m/596c4f1c-f966-4512-b9c9-dd8fbcaf0928%40postgrespro.ru
Author: Alexey Makhmutov <a.makhmutov@postgrespro.ru>
Reviewed-by: Andrey Borodin <x4mmm@yandex-team.ru>
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Alexander Korotkov <aekorotkov@gmail.com>
