Commit 36a35c550a turned the interface between ginPlaceToPage and
its subroutines in gindatapage.c and ginentrypage.c into a royal mess:
page-update critical sections were started in one place and finished in
another place not even in the same file, and the very same subroutine
might return having started a critical section or not. Subsequent patches
band-aided over some of the problems with this design by making things
even messier.
One user-visible resulting problem is memory leaks caused by the need for
the subroutines to allocate storage that would survive until ginPlaceToPage
calls XLogInsert (as reported by Julien Rouhaud). This would not typically
be noticeable during retail index updates. It could be visible in a GIN
index build, in the form of memory consumption swelling to several times
the commanded maintenance_work_mem.
Another rather nasty problem is that in the internal-page-splitting code
path, we would clear the child page's GIN_INCOMPLETE_SPLIT flag well before
entering the critical section that it's supposed to be cleared in; a
failure in between would leave the index in a corrupt state. There were
also assorted coding-rule violations with little immediate consequence but
possible long-term hazards, such as beginning an XLogInsert sequence before
entering a critical section, or calling elog(DEBUG) inside a critical
section.
To fix, redefine the API between ginPlaceToPage() and its subroutines
by splitting the subroutines into two parts. The "beginPlaceToPage"
subroutine does what can be done outside a critical section, including
full computation of the result pages into temporary storage when we're
going to split the target page. The "execPlaceToPage" subroutine is called
within a critical section established by ginPlaceToPage(), and it handles
the actual page update in the non-split code path. The critical section,
as well as the XLOG insertion call sequence, are both now always started
and finished in ginPlaceToPage(). Also, make ginPlaceToPage() create and
work in a short-lived memory context to eliminate the leakage problem.
(Since a short-lived memory context had been getting created in the most
common code path in the subroutines, this shouldn't cause any noticeable
performance penalty; we're just moving the overhead up one call level.)
In passing, fix a bunch of comments that had gone unmaintained throughout
all this klugery.
Report: <571276DD.5050303@dalibo.com>
Per discussion, there doesn't seem to be much value in having
NUM_SPINLOCK_SEMAPHORES set to 1024: under any scenario where you are
running more than a few backends concurrently, you really had better have a
real spinlock implementation if you want tolerable performance. And 1024
semaphores is a sizable fraction of the system-wide SysV semaphore limit
on many platforms. Therefore, reduce this setting's default value to 128
to make it less likely to cause out-of-semaphores problems.
Careless coding added by commit 07cacba983 could result in a crash
or a bizarre error message if someone tried to select an index on the
OID column as the replica identity index for a table. Back-patch to 9.4
where the feature was introduced.
Discussion: CAKJS1f8TQYgTRDyF1_u9PVCKWRWz+DkieH=U7954HeHVPJKaKg@mail.gmail.com
David Rowley
The code had a query-lifespan memory leak when encountering GIN entries
that have posting lists (rather than posting trees, ie, there are a
relatively small number of heap tuples containing this index key value).
With a suitable data distribution this could add up to a lot of leakage.
Problem seems to have been introduced by commit 36a35c550, so back-patch
to 9.4.
Julien Rouhaud
When re-reading an update involving both an old tuple and a new tuple from
disk, reorderbuffer.c was careless about whether the new tuple is suitably
aligned for direct access --- in general, it isn't. We'd missed seeing
this in the buildfarm because the contrib/test_decoding tests exercise this
code path only a few times, and by chance all of those cases have old
tuples with length a multiple of 4, which is usually enough to make the
access to the new tuple's t_len safe. For some still-not-entirely-clear
reason, however, Debian's sparc build gets a bus error, as reported by
Christoph Berg; perhaps it's assuming 8-byte alignment of the pointer?
The lack of previous field reports is probably because you need all of
these conditions to trigger a crash: an alignment-picky platform (not
Intel), a transaction large enough to spill to disk, an update within
that xact that changes a primary-key field and has an odd-length old tuple,
and of course logical decoding tracing the transaction.
Avoid the alignment assumption by using memcpy instead of fetching t_len
directly, and add a test case that exposes the crash on picky platforms.
Back-patch to 9.4 where the bug was introduced.
Discussion: <20160413094117.GC21485@msg.credativ.de>
It was declared as "pid_t", which would be fine except that none of
the places that printed it in error messages took any thought for the
possibility that it's not equivalent to "int". This leads to warnings
on some buildfarm members, and could possibly lead to actually wrong
error messages on those platforms. There doesn't seem to be any very
good reason not to just make it "int"; it's only ever assigned from
MyProcPid, which is int. If we want to cope with PIDs that are wider
than int, this is not the place to start.
Also, fix the comment, which seems to perhaps be a leftover from a time
when the field was only a bool?
Per buildfarm. Back-patch to 9.5 which has same issue.
As reported by Michael Feld, pg_upgrade'ing an installation having
extensions with operator families that contain just a single operator class
failed to reproduce the extension membership of those operator families.
This caused no immediate ill effects, but would create problems when later
trying to do a plain dump and restore, because the seemingly-not-part-of-
the-extension operator families would appear separately in the pg_dump
output, and then would conflict with the families created by loading the
extension. This has been broken ever since extensions were introduced,
and many of the standard contrib extensions are affected, so it's a bit
astonishing nobody complained before.
The cause of the problem is a perhaps-ill-considered decision to omit
such operator families from pg_dump's output on the grounds that the
CREATE OPERATOR CLASS commands could recreate them, and having explicit
CREATE OPERATOR FAMILY commands would impede loading the dump script into
pre-8.3 servers. Whatever the merits of that decision when 8.3 was being
written, it looks like a poor tradeoff now. We can fix the pg_upgrade
problem simply by removing that code, so that the operator families are
dumped explicitly (and then will be properly made to be part of their
extensions).
Although this fixes the behavior of future pg_upgrade runs, it does nothing
to clean up existing installations that may have improperly-linked operator
families. Given the small number of complaints to date, maybe we don't
need to worry about providing an automated solution for that; anyone who
needs to clean it up can do so with manual "ALTER EXTENSION ADD OPERATOR
FAMILY" commands, or even just ignore the duplicate-opfamily errors they
get during a pg_restore. In any case we need this fix.
Back-patch to all supported branches.
Discussion: <20228.1460575691@sss.pgh.pa.us>
When IF NOT EXISTS was added to CREATE TABLE AS, this logic didn't get
the memo, possibly resulting in an Assert failure. It looks like there
would have been no ill effects in a non-Assert build, though. Back-patch
to 9.5 where the IF NOT EXISTS option was added.
Stas Kelvich
It turns out that those PyErr_Clear() calls I removed from plpy_elog.c
in 7e3bb08038 et al were not quite as random as they appeared: they
mask a Python 2.3.x bug. (Specifically, it turns out that PyType_Ready()
can fail if the error indicator is set on entry, and PLy_traceback's fetch
of frame.f_code may be the first operation in a session that requires the
"frame" type to be readied. Ick.) Put back the clear call, but in a more
centralized place closer to what it's protecting, and this time with a
comment warning what it's really for.
Per buildfarm member prairiedog. Although prairiedog was only failing
on HEAD, it seems clearly possible for this to occur in older branches
as well, so back-patch to 9.2 the same as the previous patch.
PLy_elog() could attempt to access strings that Python had already freed,
because the strings that PLy_get_spi_error_data() returns are simply
pointers into storage associated with the error "val" PyObject. That's
fine at the instant PLy_get_spi_error_data() returns them, but just after
that PLy_traceback() intentionally releases the only refcount on that
object, allowing it to be freed --- so that the strings we pass to
ereport() are dangling pointers.
In principle this could result in garbage output or a coredump. In
practice, I think the risk is pretty low, because there are no Python
operations between where we decrement that refcount and where we use the
strings (and copy them into PG storage), and thus no reason for Python
to recycle the storage. Still, it's clearly hazardous, and it leads to
Valgrind complaints when running under a Valgrind that hasn't been
lobotomized to ignore Python memory allocations.
The code was a mess anyway: we fetched the error data out of Python
(clearing Python's error indicator) with PyErr_Fetch, examined it, pushed
it back into Python with PyErr_Restore (re-setting the error indicator),
then immediately pulled it back out with another PyErr_Fetch. Just to
confuse matters even more, there were some gratuitous-and-yet-hazardous
PyErr_Clear calls in the "examine" step, and we didn't get around to doing
PyErr_NormalizeException until after the second PyErr_Fetch, making it even
less clear which object was being manipulated where and whether we still
had a refcount on it. (If PyErr_NormalizeException did substitute a
different "val" object, it's possible that the problem could manifest for
real, because then we'd be doing assorted Python stuff with no refcount
on the object we have string pointers into.)
So, rearrange all that into some semblance of sanity, and don't decrement
the refcount on the Python error objects until the end of PLy_elog().
In HEAD, I failed to resist the temptation to reformat some messy bits
from 5c3c3cd0a3 along the way.
Back-patch as far as 9.2, because the code is substantially the same
that far back. I believe that 9.1 has the bug as well; but the code
around it is rather different and I don't want to take a chance on
breaking something for what seems a low-probability problem.
Most of what is produced by the detailed verbosity level is of no
interest at all, so switch to the normal level for more usable output.
Christian Ullrich
Backpatch to all live branches
Don't try to examine S_ISLNK(st.st_mode) after a failed lstat().
It's undefined.
Also, if the lstat() reported ENOENT, we do not wish that to be a hard
error, but the code might nonetheless treat it as one (giving an entirely
misleading error message, too) depending on luck-of-the-draw as to what
S_ISLNK() returned.
Don't throw error for ENOENT from rmdir(), either. (We're not really
expecting ENOENT because we just stat'd the file successfully; but
if we're going to allow ENOENT in the symlink code path, surely the
directory code path should too.)
Generate an appropriate errcode for its-the-wrong-type-of-file complaints.
(ERRCODE_SYSTEM_ERROR doesn't seem appropriate, and failing to write
errcode() around it certainly doesn't work, and not writing an errcode
at all is not per project policy.)
Valgrind noticed the undefined S_ISLNK result; the other problems emerged
while reading the code in the area.
All of this appears to have been introduced in 8f15f74a44.
Back-patch to 9.5 where that commit appeared.
As noted by Julian Schauder in bug #14063, the configuration-file parser
doesn't support embedded newlines in string literals. While there might
someday be a good reason to remove that restriction, there doesn't seem
to be one right now. However, ALTER SYSTEM SET could accept strings
containing newlines, since many of the variable-specific value-checking
routines would just see a newline as whitespace. This led to writing a
postgresql.auto.conf file that was broken and had to be removed manually.
Pending a reason to work harder, just throw an error if someone tries this.
In passing, fix several places in the ALTER SYSTEM logic that failed to
provide an errcode() for an ereport(), and thus would falsely log the
failure as an internal XX000 error.
Back-patch to 9.4 where ALTER SYSTEM was introduced.
The first iteration of the signal-checking loop would compute sigmask(0)
which expands to 1<<(-1) which is undefined behavior according to the
C standard. The lack of field reports of trouble suggest that it
evaluates to 0 on all existing Windows compilers, but that's hardly
something to rely on. Since signal 0 isn't a queueable signal anyway,
we can just make the loop iterate from 1 instead, and save a few cycles
as well as avoiding the undefined behavior.
In passing, avoid evaluating the volatile expression UNBLOCKED_SIGNAL_QUEUE
twice in a row; there's no reason to waste cycles like that.
Noted by Aleksander Alekseev, though this isn't his proposed fix.
Back-patch to all supported branches.
The server hasn't paid attention to the TZ environment variable since
commit ca4af308c3, but that commit missed removing this documentation
reference, as did commit d883b916a9 which added the reference where
it now belongs (initdb).
Back-patch to 9.2 where the behavior changed. Also back-patch
d883b916a9 as needed.
Matthew Somerville
The description of what the per-transaction log file says for skipped
transactions is just plain wrong.
Report and patch by Tomas Vondra, reviewed by Fabien Coelho and
modified by me.
Whenever this function is used with the FORMAT_MESSAGE_FROM_SYSTEM flag,
it's good practice to include FORMAT_MESSAGE_IGNORE_INSERTS as well.
Otherwise, if the message contains any %n insertion markers, the function
will try to fetch argument strings to substitute --- which we are not
passing, possibly leading to a crash. This is exactly analogous to the
rule about not giving printf() a format string you're not in control of.
Noted and patched by Christian Ullrich.
Back-patch to all supported branches.
Documentation mentioned B-tree, GiST and GIN as able to do multicolumn
indexes; I failed to add BRIN to the list.
Author: Petr Jediný
Reviewed-By: Fujii Masao, Emre Hasegeli
brin_page_type() and brin_metapage_info() did not enforce being called
by superuser, like other pageinspect functions that take bytea do.
Since they don't verify the passed page thoroughly, it is possible to
use them to read the server memory with a carefully crafted bytea value,
up to a file kilobytes from where the input bytea is located.
Have them throw errors if called by a non-superuser.
Report and initial patch: Andreas Seltenreich
Security: CVE-2016-3065
In the plancache, we check if the environment we planned the query under
has changed in a way which requires us to re-plan, such as when the user
for whom the plan was prepared changes and RLS is being used (and,
therefore, there may be different policies to apply).
Unfortunately, while those values were set and checked, they were not
being reset when the query was re-planned and therefore, in cases where
we change role, re-plan, and then change role again, we weren't
re-planning again. This leads to potentially incorrect policies being
applied in cases where role-specific policies are used and a given query
is planned under one role and then executed under other roles, which
could happen under security definer functions or when a common user and
query is planned initially and then re-used across multiple SET ROLEs.
Further, extensions which made use of CopyCachedPlan() may suffer from
similar issues as the RLS-related fields were not properly copied as
part of the plan and therefore RevalidateCachedQuery() would copy in the
current settings without invalidating the query.
Fix by using the same approach used for 'search_path', where we set the
correct values in CompleteCachedPlan(), check them early on in
RevalidateCachedQuery() and then properly reset them if re-planning.
Also, copy through the values during CopyCachedPlan().
Pointed out by Ashutosh Bapat. Reviewed by Michael Paquier.
Back-patch to 9.5 where RLS was introduced.
Security: CVE-2016-2193
Previously pg_rewind did not fsync any files. That's problematic, given
that the target directory is modified. If the database was started
afterwards, 2ce439f33 luckily already caused the data directory to be
synced to disk at postmaster startup; reducing the scope of the problem.
To fix, use initdb -S, at the end of the pg_rewind run. It doesn't seem
worthwhile to duplicate the code into pg_rewind, and initdb -S is
already used that way by pg_upgrade.
Reported-By: Andres Freund
Author: Michael Paquier, somewhat edited by me
Discussion: 20160310034352.iuqgvpmg5qmnxtkz@alap3.anarazel.deCAB7nPqSytVG1o4S3S2pA1O=692ekurJ+fckW2PywEG3sNw54Ow@mail.gmail.com
Backpatch: 9.5, where pg_rewind was introduced
This was a relatively harmless leak, as createBackupLabel() is only
called once per pg_rewind invocation.
Author: Michael Paquier
Reported-By: Michael Paquier
Discussion: CAB7nPqRnOw30gOXe2_SPLjh37bgm4V+txbYAPwoXb97nGQ297w@mail.gmail.com
Backpatch: 9.5, where pg_rewind was introduced
Returning the direct result of bit arithmetic, in a macro intended to be
used in a boolean manner, can be problematic if the return value is
stored in a variable of type 'bool'. If bool is implemented using C99's
_Bool, that can lead to comparison failures if the variable is then
compared again with the expression (see ginStepRight() for an example
that fails), as _Bool forces the result to be 0/1. That happens in some
configurations of newer MSVC compilers. It's also problematic when
storing the result of such an expression in a narrower type.
Several gin macros have been declared in that style since gin's initial
commit in 8a3631f8d8.
There's a lot more macros like this, but this is the only one causing
regression test failures; and I don't want to commit and backpatch a
larger patch with lots of conflicts just before the next set of minor
releases.
Discussion: 20150811154237.GD17575@awork2.anarazel.de
Backpatch: All supported branches
We really need to sync all of our IANA-derived timezone code with upstream,
but that's going to be a large patch and I certainly don't care to shove
such a thing into stable branches immediately before a release. As a
stopgap, copy just the tzcode2016c logic that checks validity of timezone
abbreviations. This prevents getting multiple "time zone abbreviation
differs from POSIX standard" bleats with tzdata 2014b and later.
DST law changes in Azerbaijan, Chile, Haiti, Palestine, and Russia (Altai,
Astrakhan, Kirov, Sakhalin, Ulyanovsk regions). Historical corrections
for Lithuania, Moldova, Russia (Kaliningrad, Samara, Volgograd).
As of 2015b, the keepers of the IANA timezone database started to use
numeric time zone abbreviations (e.g., "+04") instead of inventing
abbreviations not found in the wild like "ASTT". This causes our rather
old copy of zic to whine "warning: time zone abbreviation differs from
POSIX standard" several times during "make install". This warning is
harmless according to the IANA folk, and I don't see any problems with
these abbreviations in some simple tests; but it seems like now would be
a good time to update our copy of the tzcode stuff. I'll look into that
soon.
Unfortunately, every version of glibc thus far tested has bugs whereby
strcoll() ordering does not match strxfrm() ordering as required by
the standard. This can result in, for example, corrupted indexes.
Disabling abbreviated keys in these cases slows down non-C-collation
string sorting considerably, but there seems to be no practical
alternative. Users who are confident that their libc implementations
are solid in this regard can re-enable the optimization by compiling
with TRUST_STRXFRM.
Users who have built indexes using PostgreSQL 9.5 or PostgreSQL 9.5.1
should REINDEX if there is a possibility that they may have been
affected by this problem.
Report by Marc-Olaf Jaschke. Investigation mostly by Tom Lane, with
help from Peter Geoghegan, Noah Misch, Stephen Frost, and me. Patch
by me, reviewed by Peter Geoghegan and Tom Lane.
User-facing (even tested by regression tests) error conditions were thrown
with elog(), hence had wrong SQLSTATE and were untranslatable. And the
error message texts weren't up to project style, either.
jsonb_set() could produce wrong answers or incorrect error reports, or in
the worst case even crash, when trying to convert a path-array element into
an integer for use as an array subscript. Per report from Vitaly Burovoy.
Back-patch to 9.5 where the faulty code was introduced (in commit
c6947010ce).
Michael Paquier
In commit afb9249d06, we (probably I) made ExecLockRows assign
null test tuples to all relations of the query while setting up to do an
EvalPlanQual recheck for a newly-updated locked row. This was sheerest
brain fade: we should only set test tuples for relations that are lockable
by the LockRows node, and in particular empty test tuples are only sensible
for inheritance child relations that weren't the source of the current
tuple from their inheritance tree. Setting a null test tuple for an
unrelated table causes it to return NULLs when it should not, as exhibited
in bug #14034 from Bronislav Houdek. To add insult to injury, doing it the
wrong way required two loops where one would suffice; so the corrected code
is even a bit shorter and faster.
Add a regression test case based on his example, and back-patch to 9.5
where the bug was introduced.
Modern Perl has removed psed from its core distribution, so it might not
be readily available on some build platforms. We therefore replace its
use with a Perl script generated by s2p, which is equivalent to the sed
script. The latter is retained for non-MSVC builds to avoid creating a
new hard dependency on Perl for non-Windows tarball builds.
Backpatch to all live branches.
Michael Paquier and me.
In HEAD, fix incorrect field width for hours part of OF when tm_gmtoff is
negative. This was introduced by commit 2d87eedc1d as a result of
falsely applying a pattern that's correct when + signs are omitted, which
is not the case for OF.
In 9.4, fix missing abs() call that allowed a sign to be attached to the
minutes part of OF. This was fixed in 9.5 by 9b43d73b3f, but for
inscrutable reasons not back-patched.
In all three versions, ensure that the sign of tm_gmtoff is correctly
reported even when the GMT offset is less than 1 hour.
Add regression tests, which evidently we desperately need here.
Thomas Munro and Tom Lane, per report from David Fetter
This didn't work because when we dropped and re-established a database
connection, we did not bother to reset session-specific state such as
the statements-are-prepared flags.
The st->prepared[] array certainly needs to be flushed, and I cleared a
couple of other fields as well that couldn't possibly retain meaningful
state for a new connection.
In passing, fix some bogus comments and strange field order choices.
Per report from Robins Tharakan.
INSERT ... ON CONFLICT's precheck may have to wait on the outcome of
another insertion, which may or may not itself be a speculative
insertion. This wait is not necessarily associated with an exclusion
constraint, but was always reported that way in log messages if the wait
happened to involve a tuple that had no speculative token.
Initially discovered through use of ON CONFLICT DO NOTHING, where
spurious references to exclusion constraints in log messages were more
likely.
Patch by Peter Geoghegan.
Reviewed by Julien Rouhaud.
Back-patch to 9.5 where INSERT ... ON CONFLICT was added.
Previously, we included <xlocale.h> only if necessary to get the definition
of type locale_t. According to notes in PGAC_TYPE_LOCALE_T, this is
important because on some versions of glibc that file supplies an
incompatible declaration of locale_t. (This info may be obsolete, because
on my RHEL6 box that seems to be the *only* definition of locale_t; but
there may still be glibc's in the wild for which it's a live concern.)
It turns out though that on FreeBSD and maybe other BSDen, you can get
locale_t from stdlib.h or locale.h but mbstowcs_l() and friends only from
<xlocale.h>. This was leaving us compiling calls to mbstowcs_l() and
friends with no visible prototype, which causes a warning and could
possibly cause actual trouble, since it's not declared to return int.
Hence, adjust the configure checks so that we'll include <xlocale.h>
either if it's necessary to get type locale_t or if it's necessary to
get a declaration of mbstowcs_l().
Report and patch by Aleksander Alekseev, somewhat whacked around by me.
Back-patch to all supported branches, since we have been using
mbstowcs_l() since 9.1.
On the machines I tried this on, pressing TAB after SECURITY LABEL led to
being offered ON and FOR as intended, plus random other keywords (varying
across machines). But if you were a bit more unlucky you'd get a crash,
as reported by nummervet@mail.ru in bug #14019.
Seems to have been an aboriginal error in the SECURITY LABEL patch,
commit 4d355a8336. Hence, back-patch to all supported versions.
There's no bug in HEAD, though, thanks to our recent tab-completion
rewrite.
Commit d88976cfa1 removed this code from ginFreeScanKeys():
- if (entry->list)
- pfree(entry->list);
evidently in the belief that that ItemPointer array is allocated in the
keyCtx and so would be reclaimed by the following MemoryContextReset.
Unfortunately, it isn't and it won't. It'd likely be a good idea for
that to become so, but as a simple and back-patchable fix in the
meantime, restore this code to ginFreeScanKeys().
Also, add a similar pfree to where startScanEntry() is about to zero out
entry->list. I am not sure if there are any code paths where this
change prevents a leak today, but it seems like cheap future-proofing.
In passing, make the initial allocation of so->entries[] use palloc
not palloc0. The code doesn't depend on unused entries being zero;
if it did, the array-enlargement code in ginFillScanEntry() would be
wrong. So using palloc0 initially can only serve to confuse readers
about what the invariant is.
Per report from Felipe de Jesús Molina Bravo, via Jaime Casanova in
<CAJGNTeMR1ndMU2Thpr8GPDUfiHTV7idELJRFusA5UXUGY1y-eA@mail.gmail.com>
Tom Lane