ECPG assumes that any bytea data it receives from a backend starts with
'\x' as its first two bytes, but a check was missed to enforce that. A
rogue server sending some garbage bytea data would be able to crash a
client, resulting in a client-side DoS, in the most common cases.
Reported-by: ylwangtju <ylwangtju@qq.com>
Backpatch-through: 14
Security: CVE-2026-16241
Commit f6f0542266 changed getNotify(),
getParameterStatus(), and related libpq message-processing paths to
abandon the connection on out-of-memory errors.
However, getCopyDataMessage() and pqFunctionCall3() did not handle
this new fatal-error state. Both can process asynchronous
NotificationResponse and ParameterStatus messages while waiting for
other responses. If one of those messages triggered a fatal error, these
loops continued processing instead of reporting it immediately.
Fix this by checking for a saved fatal error after processing an
asynchronous message. If the connection has been abandoned, return the
appropriate error immediately instead of continuing to parse input.
Backpatch to v18, where commit f6f0542266 introduced this issue.
Author: Anthonin Bonnefoy <anthonin.bonnefoy@datadoghq.com>
Reviewed-by: Ewan Young <kdbase.hack@gmail.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/CAO6_XqpGfm+XHE1OzS=_+jroeDOxhhGa11P3cbm9q2gT05yorA@mail.gmail.com
Backpatch-through: 18
The slow_down interval parsing code checks explicitly for overflow, but
since it does that after the signed overflow has already occurred, we
end up inviting undefined behavior from the compiler anyway.
Use checked arithmetic instead. set_timer() takes a long int in order to
interface nicely with libcurl, so use an int32 as the interval counter
and clamp to LONG_MAX during conversion to milliseconds.
Backpatch to 18, where libpq-oauth was introduced.
Reported-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Discussion: https://postgr.es/m/qtclihmrkq67ach3xjxyi4qcksstin5qxwsnkqefkmotxwh4g6%40ae2bj6jvcmry
Backpatch-through: 18
(This is a late cherry-pick of 4e1e41733, now that the new suite has
proven its stability, to ensure coverage for this code in PG18 as the
later branches diverge.)
To better record the internal behaviors of oauth-curl.c, add a unit test
suite for the socket and timer handling code. This is all based on TAP
and driven by our existing Test::More infrastructure.
This commit is a replay of 1443b6c0e, which was reverted due to
buildfarm failures. Compared with that, this version protects the build
targets in the Makefile with a with_libcurl conditional, and it tweaks
the code style in 001_oauth.pl.
Reviewed-by: Dagfinn Ilmari Mannsåker <ilmari@ilmari.org>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Discussion: https://postgr.es/m/CAOYmi+nDZxJHaWj9_jRSyf8uMToCADAmOfJEggsKW-kY7aUwHA@mail.gmail.com
Discussion: https://postgr.es/m/CAOYmi+m=xY0P_uAzAP_884uF-GhQ3wrineGwc9AEnb6fYxVqVQ@mail.gmail.com
Check explicitly for pqsecure_read() returning an error. It shouldn't
fail, and we would've caught it in the check for a short read, but
better to be explicit so that the error message is more informative.
We also shouldn't update 'inEnd' when the read fails, although that
too is just pro forma as we will bail out and close the connection on
error.
Reported-by: Peter Eisentraut <peter@eisentraut.org>
Discussion: https://www.postgresql.org/message-id/34844e8c-267c-4daf-b1e0-f26059a4a7d3@eisentraut.org
Backpatch-through: 14
The previous commit strengthened a workaround for a hang when large
messages are split across TLS records/GSS tokens. Because that
workaround is implemented in libpq internals, it can only help us when
libpq itself is polling on the socket. In nonblocking situations,
where the client above libpq is expected to poll, the same bugs can
show up.
As a contrived example, consider a large protocol-2.0 error coming
back from a server during PQconnectPoll(), split in an odd way across
two records:
-- TLS record (8192-byte payload) --
EEEE[...repeated a total of 8192 times]
-- TLS record (8193-byte payload) --
EEEE[...repeated a total of 8192 times]\0
The first record will fill the first half of the libpq receive buffer,
which is 16k long by default. The second record completely fills the
last half with its first 8192 bytes, leaving the terminating NULL in
the OpenSSL buffer. Since we still haven't seen the terminator at our
level, PQconnectPoll() will return PGRES_POLLING_READING, expecting to
come back when the server has sent "the rest" of the data. But there
is nothing left to read from the socket; OpenSSL had to pull all of
the data in the 8193-byte record off of the wire to decrypt it.
A real server would probably not split up the records this way, nor
keep the connection open after sending a fatal connection error. But
servers that regularly use larger TLS records can get the libpq
receive buffer into the same state if DataRows are big enough, as
reported on the list. While the PostgreSQL server doesn't use larger
TLS records like that, other non-PostgreSQL servers that implement the
wire protocol are known to do that, as well as proxies that sit
between the server and the client
This is a layering violation. libpq makes decisions based on data in
the application buffer, above the transport buffer (whether SSL or
GSS), but clients are polling the socket below the transport buffer.
One way to fix this in a backportable way, without changing APIs too
much, is to ensure data never stays in the transport buffer. Then
pqReadData's postconditions will look similar for both raw sockets and
SSL/GSS: any available data is either in the application buffer, or
still on the socket.
Building on the prior commit, make pqReadData() to drain all pending
data from the transport layer into conn->inBuffer, expanding the
buffer as necessary. This is not particularly efficient from an
architectural perspective (the pqsecure_read() implementations take
care to fit their packets into the current buffer, and that effort is
now completely discarded), but it's hopefully easier to reason about
than a full rewrite would be for the back branches.
Author: Jacob Champion <jacob.champion@enterprisedb.com>
Reviewed-by: Mark Dilger <mark.dilger@enterprisedb.com>
Reviewed-by: solai v <solai.cdac@gmail.com>
Reported-by: Lars Kanis <lars@greiz-reinsdorf.de>
Discussion: https://postgr.es/m/2039ac58-d3e0-434b-ac1a-2a987f3b4cb1%40greiz-reinsdorf.de
Backpatch-through: 14
An extra check for pending bytes in the SSL layer has been part of
pqReadReady() for a very long time (79ff2e96d). But when GSS transport
encryption was added, it didn't receive the same treatment. (As
79ff2e96d notes, "The bug that I fixed in this patch is exceptionally
hard to reproduce reliably.")
Without that check, it's possible to hit a hang in gssencmode, if the
server splits a large libpq message such that the final message in a
streamed response is part of the same wrapped token as the split
message:
DataRowDataRowDataRowDataRowDataRowData
-- token boundary --
RowDataRowCommandCompleteReadyForQuery
If the split message takes up enough memory to nearly fill libpq's
receive buffer, libpq may return from pqReadData() before the later
messages are pulled out of the PqGSSRecvBuffer. Without additional
socket activity from the server, pqReadReady() (via pqSocketCheck())
will never again return true, hanging the connection.
Pull the pending-bytes check into the pqsecure API layer, where both
SSL and GSS now implement it.
Note that this does not fix the root problem! Third party clients of
libpq have no way to call pqsecure_read_is_pending() in their own
polling. This just brings the GSS implementation up to par with the
existing SSL workaround; a broader fix is left to a subsequent commit.
In preparation for the broader fix, this patch already changes the
*_read_pending() functions to return the number of bytes in the buffer
rather than just a boolean. The current callers don't need that, but
the subsequent fix will.
Author: Jacob Champion <jacob.champion@enterprisedb.com>
Discussion: https://postgr.es/m/CAOYmi%2BmpymrgZ76Jre2dx_PwRniS9YZojwH0rZnTuiGHCsj0rA%40mail.gmail.com
Backpatch-through: 14
When compiling a DECLARE section containing a union nested
inside a struct, ecpg passes a null value for struct_sizeof to
ECPGmake_struct_type. I (tgl) didn't foresee that case in
commit 0e6060790, and wrote an unprotected mm_strdup() call.
Reported-by: iMSA (via Jehan-Guillaume de Rorthais <jgdr@dalibo.com>)
Author: Jehan-Guillaume de Rorthais <jgdr@dalibo.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/20260625114849.34b2148e@karst
Backpatch-through: 18
When debugging an OAuth trace, it's helpful to know what version of Curl
is in use. The SSL library that Curl is using (which may not be the one
in use by libpq) is also relevant, and it's just as easy to get, so
print that too.
This is being added post-feature-freeze, with RMT approval, in order to
fix some tests in the face of an upstream Curl regression. A subsequent
commit will make use of it in oauth_validator. Backpatch to 18 as well.
Tested-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CAOYmi%2B%3DkP86t%2BZFFXNQ9G6K4ht7utdmB%3DCzhP%3DZ2wvuBymOTtQ%40mail.gmail.com
Backpatch-through: 18
OpenSSL 4.0.0 changed some parameters and returnvalues to const, so
we need to update our declarations and subsequently cast away const-
ness from a few callsites to make libpq build without warnings. This
is tested with OpenSSL 1.1.1 through 4.0.0 as well as with LibreSSL.
No functional change is introduced, this commit only allows postgres
to be compiled against OpenSSL 4.0.0 without warnings.
There is also an errormessage change in OpenSSL 4.0.0 which needed
to be covered by our testharness.
This will be backpatched to all supported branches since they are
all equally likely to be built against OpenSSL 4.0.0 as it becomes
available in distributions. Backpatching will be done once it has
been in master for a few days without issues.
Author: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/066B07BB-85FA-487C-BE8C-40F791CFC3C4@yesql.se
Backpatch-through: 14
Previously, ecpg accepted multiple descriptor header items in GET DESCRIPTOR
and SET DESCRIPTOR, but generated broken C code when they were used.
Although the grammar allowed this syntax, the implementation did not actually
support it.
This commit tightens the ecpg grammar so the header form of GET/SET DESCRIPTOR
accepts only a single header item, matching the implementation and preventing
generation of broken C code.
Also update the documentation synopsis accordingly.
Backpatch to all supported versions.
Author: Masashi Kamura <kamura.masashi@fujitsu.com>
Reviewed-by: Hayato Kuroda <kuroda.hayato@fujitsu.com>
Reviewed-by: Lakshmi G <lakshmigcdac@gmail.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/OS9PR01MB13174AD7D1829D0644B6BB90E9447A@OS9PR01MB13174.jpnprd01.prod.outlook.com
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
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
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
ECPGdeallocate_all(), ECPGprepared_statement(), ECPGget_desc(), and
ecpg_freeStmtCacheEntry() could crash with a SIGSEGV when called
without an established connection (for example, when EXEC SQL CONNECT
was forgotten or a non-existent connection name was used), because
they dereferenced the result of ecpg_get_connection() without first
checking it for NULL.
Each site is fixed in the style of the surrounding code.
New tests are added for these conditions.
Author: Shruthi Gowda <gowdashru@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Mahendra Singh Thalor <mahi6run@gmail.com>
Reviewed-by: Nishant Sharma <nishant.sharma@enterprisedb.com>
Discussion: https://postgr.es/m/3007317.1765210195@sss.pgh.pa.us
Backpatch-through: 14
Newest versions of gcc+glibc are able to detect cases where code
implicitly casts away const by assigning the result of strchr() or
a similar function applied to a "const char *" value to a target
variable that's just "char *". This of course creates a hazard of
not getting a compiler warning about scribbling on a string one was
not supposed to, so fixing up such cases is good.
This patch fixes a dozen or so places where we were doing that.
Most are trivial additions of "const" to the target variable,
since no actually-hazardous change was occurring.
Thanks to Bertrand Drouvot for finding a couple more spots than
I had.
This commit back-patches relevant portions of 8f1791c61 and
9f7565c6c into supported branches. However, there are two
places in ecpg (in v18 only) where a proper fix is more
complicated than seems appropriate for a back-patch. I opted
to silence those two warnings by adding casts.
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Discussion: https://postgr.es/m/1324889.1764886170@sss.pgh.pa.us
Discussion: https://postgr.es/m/3988414.1771950285@sss.pgh.pa.us
Backpatch-through: 14-18
The definition of PGoauthBearerRequest uses a temporary SOCKTYPE macro
to hide the difference between Windows and Berkeley socket handles,
since we don't surface pgsocket in our public API. This macro doesn't
need to escape the header, because implementers will choose the correct
socket type based on their platform, so I #undef'd it immediately after
use.
I didn't namespace that helper, though, so if anyone else needs a
SOCKTYPE macro, libpq-fe.h will now unhelpfully get rid of it. This
doesn't seem too far-fetched, given its proximity to existing POSIX
macro names.
Add a PQ_ prefix to avoid collisions, update and improve the surrounding
documentation, and backpatch.
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Discussion: https://postgr.es/m/CAOYmi%2BmrGg%2Bn_X2MOLgeWcj3v_M00gR8uz_D7mM8z%3DdX1JYVbg%40mail.gmail.com
Backpatch-through: 18
Now that the prior commits have fixed missing OAuth translations, pull
the bespoke usage of libpq_gettext() for OAUTHBEARER parsing into
oauth_json_set_error() itself, and make that a gettext trigger as well,
to better match what the other sites are doing. Add an _internal()
variant to handle the existing untranslated case.
Suggested-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Discussion: https://postgr.es/m/0EEBCAA8-A5AC-4E3B-BABA-0BA7A08C361B%40yesql.se
Backpatch-through: 18
Some error messages are generated when OAuth multiplexer operations fail
unexpectedly in the client. Álvaro pointed out that these are both
difficult to translate idiomatically (as they use internal terminology
heavily) and of dubious translation value to end users (since they're
going to need to get developer help anyway). The response parsing engine
has a similar issue.
Remove these from the translation files by introducing internal variants
of actx_error() and oauth_parse_set_error().
Suggested-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Discussion: https://postgr.es/m/CAOYmi%2BkQQ8vpRcoSrA5EQ98Wa3G6jFj1yRHs6mh1V7ohkTC7JA%40mail.gmail.com
Backpatch-through: 18
Several strings that should have been translated as they passed through
libpq_gettext were not actually being pulled into the translation files,
because I hadn't directly wrapped them in one of the GETTEXT_TRIGGERS.
Move the responsibility for calling libpq_gettext() to the code that
sets actx->errctx. Doing the same in report_type_mismatch() would result
in double-translation, so mark those strings with gettext_noop()
instead. And wrap two ternary operands with gettext_noop(), even though
they're already in one of the triggers, since xgettext sees only the
first.
Finally, fe-auth-oauth.c was missing from nls.mk, so none of that file
was being translated at all. Add it now.
Original patch by Zhijie Hou, plus suggested tweaks by Álvaro Herrera
and small additions by me.
Reported-by: Zhijie Hou <houzj.fnst@fujitsu.com>
Author: Zhijie Hou <houzj.fnst@fujitsu.com>
Co-authored-by: Álvaro Herrera <alvherre@kurilemu.de>
Co-authored-by: Jacob Champion <jacob.champion@enterprisedb.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Discussion: https://postgr.es/m/TY4PR01MB1690746DB91991D1E9A47F57E94CDA%40TY4PR01MB16907.jpnprd01.prod.outlook.com
Backpatch-through: 18
Instead of having to write a semicolon inside the macro argument, we can
insert a semicolon with another macro layer. This no longer gives
pg_bsd_indent indigestion, so we can remove the digestive aids that had
to be installed in the pgindent Perl script.
Author: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/202511111134.njrwf5w5nbjm@alvherre.pgsql
Backpatch-through: 18
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
CRLF translation is already handled by the text mode, so there should be
no need for any specific logic. See also 1c6d462939, msys perl being
one case where the translation mattered.
Note: An equivalent has been first applied on HEAD with 8767b449a3.
The same change is backpatched to v18 as all the Windows buildfarm
members have reported green.
Author: Jacob Champion <jacob.champion@enterprisedb.com>
Co-authored-by: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Discussion: https://postgr.es/m/aPsh39bxwYKvUlAf@paquier.xyz
Backpatch-through: 18
On Windows, this code did not handle error conditions correctly at
all, since it looked at "errno" which is not used for socket-related
errors on that platform. This resulted, for example, in failure
to connect to a PostgreSQL server with GSSAPI enabled.
We have a convention for dealing with this within libpq, which is to
use SOCK_ERRNO and SOCK_ERRNO_SET rather than touching errno directly;
but the GSSAPI code is a relative latecomer and did not get that memo.
(The equivalent backend code continues to use errno, because the
backend does this differently. Maybe libpq's approach should be
rethought someday.)
Apparently nobody tries to build libpq with GSSAPI support on Windows,
or we'd have heard about this before, because it's been broken all
along. Back-patch to all supported branches.
Author: Ning Wu <ning94803@gmail.com>
Co-authored-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CAFGqpvg-pRw=cdsUpKYfwY6D3d-m9tw8WMcAEE7HHWfm-oYWvw@mail.gmail.com
Backpatch-through: 13
Fix for commit 3c86223c99. That commit moved the typedef of pg_int64
from postgres_ext.h to libpq-fe.h, because the only remaining place
where it might be used is libpq users, and since the type is obsolete,
the intent was to limit its scope.
The problem is that if someone builds an extension against an
older (pre-PG18) server version and a new (PG18) libpq, they might get
two typedefs, depending on include file order. This is not allowed
under C99, so they might get warnings or errors, depending on the
compiler and options. The underlying types might also be
different (e.g., long int vs. long long int), which would also lead to
errors. This scenario is plausible when using the standard Debian
packaging, which provides only the newest libpq but per-major-version
server packages.
The fix is to undo that part of commit 3c86223c99. That way, the
typedef is in the same header file across versions. At least, this is
the safest fix doable before PostgreSQL 18 releases.
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://www.postgresql.org/message-id/25144219-5142-4589-89f8-4e76948b32db%40eisentraut.org
PQtrace() was generating its output for non-printable characters without
casting the characters printed with unsigned char, leading to some extra
"\xffffff" generated in the output due to the fact that char may be
signed.
Oversights introduced by commit 198b3716db, so backpatch down to v14.
Author: Ran Benita <ran@unusedvar.com>
Discussion: https://postgr.es/m/a3383211-4539-459b-9d51-95c736ef08e0@app.fastmail.com
Backpatch-through: 14
libpq-oauth uses floor() but did not link against libm. Since libpq
itself uses -lm, nothing in the buildfarm has had problems with
libpq-oauth yet, and it seems difficult to hit a failure in practice.
But commit 1443b6c0e attempted to add an executable based on
libpq-oauth, which ran into link-time failures with Clang due to this
omission. It seems prudent to fix this for both the module and the
executable simultaneously so that no one trips over it in the future.
This is a Makefile-only change. The Meson side already pulls in libm,
through the os_deps dependency.
Discussion: https://postgr.es/m/CAOYmi%2Bn6ORcmV10k%2BdAs%2Bp0b9QJ4bfpk0WuHQaF5ODXxM8Y36A%40mail.gmail.com
Backpatch-through: 18
The protocol documentation states that the maximum length of a cancel
key is 256 bytes. This starts checking for that limit in libpq.
Otherwise third party backend implementations will probably start
using more bytes anyway. We also start requiring that a protocol 3.0
connection does not send a longer cancel key, to make sure that
servers don't start breaking old 3.0-only clients by accident. Finally
this also restricts the minimum key length to 4 bytes (both in the
protocol spec and in the libpq implementation).
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Jacob Champion <jchampion@postgresql.org>
Discussion: https://www.postgresql.org/message-id/df892f9f-5923-4046-9d6f-8c48d8980b50@iki.fi
Backpatch-through: 18
In most cases, if an out-of-memory situation happens, we attach the
error message to the connection and report it at the next
PQgetResult() call. However, there are a few cases, while processing
messages that are not associated with any particular query, where we
handled failed allocations differently and not very nicely:
- If we ran out of memory while processing an async notification,
getNotify() either returned EOF, which stopped processing any
further data until more data was received from the server, or
silently dropped the notification. Returning EOF is problematic
because if more data never arrives, e.g. because the connection was
used just to wait for the notification, or because the next
ReadyForQuery was already received and buffered, it would get stuck
forever. Silently dropping a notification is not nice either.
- (New in v18) If we ran out of memory while receiving BackendKeyData
message, getBackendKeyData() returned EOF, which has the same issues
as in getNotify().
- If we ran out of memory while saving a received a ParameterStatus
message, we just skipped it. A later call to PQparameterStatus()
would return NULL, even though the server did send the status.
Change all those cases to terminate the connnection instead. Our
options for reporting those errors are limited, but it seems better to
terminate than try to soldier on. Applications should handle
connection loss gracefully, whereas silently missing a notification,
parameter status, or cancellation key could cause much weirder
problems.
This also changes the error message on OOM while expanding the input
buffer. It used to report "cannot allocate memory for input buffer",
followed by "lost synchronization with server: got message type ...".
The "lost synchronization" message seems unnecessary, so remove that
and report only "cannot allocate memory for input buffer". (The
comment speculated that the out of memory could indeed be caused by
loss of sync, but that seems highly unlikely.)
This evolved from a more narrow patch by Jelte Fennema-Nio, which was
reviewed by Jacob Champion.
Somewhat arbitrarily, backpatch to v18 but no further. These are
long-standing issues, but we haven't received any complaints from the
field. We can backpatch more later, if needed.
Co-authored-by: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Jacob Champion <jchampion@postgresql.org>
Discussion: https://www.postgresql.org/message-id/df892f9f-5923-4046-9d6f-8c48d8980b50@iki.fi
Backpatch-through: 18
Tracking down the bugs that led to the addition of comb_multiplexer()
and drain_timer_events() was difficult, because an inefficient flow is
not visibly different from one that is working properly. To help
maintainers notice when something has gone wrong, track the number of
calls into the flow as part of debug mode, and print the total when the
flow finishes.
A new test makes sure the total count is less than 100. (We expect
something on the order of 10.) This isn't foolproof, but it is able to
catch several regressions in the logic of the prior two commits, and
future work to add TLS support to the oauth_validator test server should
strengthen it as well.
Backpatch-through: 18
Discussion: https://postgr.es/m/CAOYmi+nDZxJHaWj9_jRSyf8uMToCADAmOfJEggsKW-kY7aUwHA@mail.gmail.com
In a case similar to the previous commit, an expired timer can remain
permanently readable if Curl does not remove the timeout itself. Since
that removal isn't guaranteed to happen in real-world situations,
implement drain_timer_events() to reset the timer before calling into
drive_request().
Moving to drain_timer_events() happens to fix a logic bug in the
previous caller of timer_expired(), which treated an error condition as
if the timer were expired instead of bailing out.
The previous implementation of timer_expired() gave differing results
for epoll and kqueue if the timer was reset. (For epoll, a reset timer
was considered to be expired, and for kqueue it was not.) This didn't
previously cause problems, since timer_expired() was only called while
the timer was known to be set, but both implementations now use the
kqueue logic.
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Backpatch-through: 18
Discussion: https://postgr.es/m/CAOYmi+nDZxJHaWj9_jRSyf8uMToCADAmOfJEggsKW-kY7aUwHA@mail.gmail.com
If Curl needs to switch the direction of a socket's registration (e.g.
from CURL_POLL_IN to CURL_POLL_OUT), it expects the old registration to
be discarded. For epoll, this happened via EPOLL_CTL_MOD, but for
kqueue, the old registration would remain if it was not explicitly
removed by Curl.
Explicitly remove the opposite-direction event during registrations. (If
that event doesn't exist, we'll just get an ENOENT, which will be
ignored by the same code that handles CURL_POLL_REMOVE.) A few
assertions are also added to strengthen the relationship between the
number of events added, the number of events pulled off the queue, and
the lengths of the kevent arrays.
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Backpatch-through: 18
Discussion: https://postgr.es/m/CAOYmi+nDZxJHaWj9_jRSyf8uMToCADAmOfJEggsKW-kY7aUwHA@mail.gmail.com
If a socket is added to the kqueue, becomes readable/writable, and
subsequently becomes non-readable/writable again, the kqueue itself will
remain readable until either the socket registration is removed, or the
stale event is cleared via a call to kevent().
In many simple cases, Curl itself will remove the socket registration
quickly, but in real-world usage, this is not guaranteed to happen. The
kqueue can then remain stuck in a permanently readable state until the
request ends, which results in pointless wakeups for the client and
wasted CPU time.
Implement comb_multiplexer() to call kevent() and unstick any stale
events that would cause unnecessary callbacks. This is called right
after drive_request(), before we return control to the client to wait.
Suggested-by: Thomas Munro <thomas.munro@gmail.com>
Co-authored-by: Thomas Munro <thomas.munro@gmail.com>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Backpatch-through: 18
Discussion: https://postgr.es/m/CAOYmi+nDZxJHaWj9_jRSyf8uMToCADAmOfJEggsKW-kY7aUwHA@mail.gmail.com
PostgreSQL always sends the BackendKeyData message at connection
startup, but there are some third party backend implementations out
there that don't support cancellation, and don't send the message
[1]. While the protocol docs left it up for interpretation if that is
valid behavior, libpq in PostgreSQL 17 and below accepted it. It does
not seem like the libpq behavior was intentional though, since it did
so by sending CancelRequest messages with all zeros to such servers
(instead of returning an error or making the cancel a no-op).
In version 18 the behavior was changed to return an error when trying
to create the cancel object with PGgetCancel() or PGcancelCreate().
This was done without any discussion, as part of supporting different
lengths of cancel packets for the new 3.2 version of the protocol.
This commit changes the behavior of PGgetCancel() / PGcancel() once
more to only return an error when the cancel object is actually used
to send a cancellation, instead of when merely creating the object.
The reason to do so is that some clients [2] create a cancel object as
part of their connection creation logic (thus having the cancel object
ready for later when they need it), so if creating the cancel object
returns an error, the whole connection attempt fails. By delaying the
error, such clients will still be able to connect to the third party
backend implementations in question, but when actually trying to
cancel a query, the user will be notified that that is not possible
for the server that they are connected to.
This commit only changes the behavior of the older PGgetCancel() /
PQcancel() functions, not the more modern PQcancelCreate() family of
functions. I.e. PQcancelCreate() returns a failed connection object
(CONNECTION_BAD) if the server didn't send a cancellation key. Unlike
the old PQgetCancel() function, we're not aware of any clients in the
field that use PQcancelCreate() during connection startup in a way
that would prevent connecting to such servers.
[1] AWS RDS Proxy is definitely one of them, and CockroachDB might be
another.
[2] psycopg2 (but not psycopg3).
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Jacob Champion <jacob.champion@enterprisedb.com>
Backpatch-through: 18
Discussion: https://www.postgresql.org/message-id/20250617.101056.1437027795118961504.ishii%40postgresql.org
