Previously, a worker process would establish values for these based on
its own start time. In v10 and up, this can trivially be shown to cause
misbehavior of transaction_timestamp(), timestamp_in(), and related
functions which are (perhaps unwisely?) marked parallel-safe. It seems
likely that other behaviors might diverge from what happens in the parent
as well.
It's not as trivial to demonstrate problems in 9.6 or 9.5, but I'm sure
it's still possible, so back-patch to all branches containing parallel
worker infrastructure.
In HEAD only, mark now() and statement_timestamp() as parallel-safe
(other affected functions already were). While in theory we could
still squeeze that change into v11, it doesn't seem important enough
to force a last-minute catversion bump.
Konstantin Knizhnik, whacked around a bit by me
Discussion: https://postgr.es/m/6406dbd2-5d37-4cb6-6eb2-9c44172c7e7c@postgrespro.ru
Historically we forbade datatype-specific comparison functions from
returning INT_MIN, so that it would be safe to invert the sort order
just by negating the comparison result. However, this was never
really safe for comparison functions that directly return the result
of memcmp(), strcmp(), etc, as POSIX doesn't place any such restriction
on those library functions. Buildfarm results show that at least on
recent Linux on s390x, memcmp() actually does return INT_MIN sometimes,
causing sort failures.
The agreed-on answer is to remove this restriction and fix relevant
call sites to not make such an assumption; code such as "res = -res"
should be replaced by "INVERT_COMPARE_RESULT(res)". The same is needed
in a few places that just directly negated the result of memcmp or
strcmp.
To help find places having this problem, I've also added a compile option
to nbtcompare.c that causes some of the commonly used comparators to
return INT_MIN/INT_MAX instead of their usual -1/+1. It'd likely be
a good idea to have at least one buildfarm member running with
"-DSTRESS_SORT_INT_MIN". That's far from a complete test of course,
but it should help to prevent fresh introductions of such bugs.
This is a longstanding portability hazard, so back-patch to all supported
branches.
Discussion: https://postgr.es/m/20180928185215.ffoq2xrq5d3pafna@alap3.anarazel.de
The API (EOH_flatten_into) that flattens the expanded value representation
expects the target address to be maxaligned. All it's usage adhere to that
principle except when serializing datums for parallel query. Fix that
usage.
Diagnosed-by: Tom Lane
Author: Tom Lane and Amit Kapila
Backpatch-through: 9.6
Discussion: https://postgr.es/m/11629.1536550032@sss.pgh.pa.us
Previously, we used the platform's NL_ARGMAX if any, otherwise 16.
The trouble with this is that the platform value is hugely variable,
ranging from the POSIX-minimum 9 to as much as 64K on recent FreeBSD.
Values of more than a dozen or two have no practical use and slow down
the initialization of the argtypes array. Worse, they cause snprintf.c
to consume far more stack space than was the design intention, possibly
resulting in stack-overflow crashes.
Standardize on 31, which is comfortably more than we need (it looks like
no existing translatable message has more than about 10 parameters).
I chose that, not 32, to make the array sizes powers of 2, for some
possible small gain in speed of the memset.
The lack of reported crashes suggests that the set of platforms we
use snprintf.c on (in released branches) may have no overlap with
the set where NL_ARGMAX has unreasonably large values. But that's
not entirely clear, so back-patch to all supported branches.
Per report from Mateusz Guzik (via Thomas Munro).
Discussion: https://postgr.es/m/CAEepm=3VF=PUp2f8gU8fgZB22yPE_KBS0+e1AHAtQ=09schTHg@mail.gmail.com
The variants of these functions that take numeric inputs (OIDs or
column numbers) are supposed to return NULL rather than failing
on bad input; this rule reduces problems with snapshot skew when
queries apply the functions to all rows of a catalog.
has_column_privilege() had careless handling of the case where the
table OID didn't exist. You might get something like this:
select has_column_privilege(9999,'nosuchcol','select');
ERROR: column "nosuchcol" of relation "(null)" does not exist
or you might get a crash, depending on the platform's printf's response
to a null string pointer.
In addition, while applying the column-number variant to a dropped
column returned NULL as desired, applying the column-name variant
did not:
select has_column_privilege('mytable','........pg.dropped.2........','select');
ERROR: column "........pg.dropped.2........" of relation "mytable" does not exist
It seems better to make this case return NULL as well.
Also, the OID-accepting variants of has_foreign_data_wrapper_privilege,
has_server_privilege, and has_tablespace_privilege didn't follow the
principle of returning NULL for nonexistent OIDs. Superusers got TRUE,
everybody else got an error.
Per investigation of Jaime Casanova's report of a new crash in HEAD.
These behaviors have been like this for a long time, so back-patch to
all supported branches.
Patch by me; thanks to Stephen Frost for discussion and review
Discussion: https://postgr.es/m/CAJGNTeP=-6Gyqq5TN9OvYEydi7Fv1oGyYj650LGTnW44oAzYCg@mail.gmail.com
If the column being modified is referenced by a foreign key constraint
of another table, ALTER TABLE would open the other table (to re-parse
the constraint's definition) without having first obtained a lock on it.
This was evidently intentional, but that doesn't mean it's really safe.
It's especially not safe in 9.3, which pre-dates use of MVCC scans for
catalog reads, but even in current releases it doesn't seem like a good
idea.
We know we'll need AccessExclusiveLock shortly to drop the obsoleted
constraint, so just get that a little sooner to close the hole.
Per testing with a patch that complains if we open a relation without
holding any lock on it. I don't plan to back-patch that patch, but we
should close the holes it identifies in all supported branches.
Discussion: https://postgr.es/m/2038.1538335244@sss.pgh.pa.us
The method we've traditionally used, of redeclaring strerror_r() to
see if the compiler complains of inconsistent declarations, turns out
not to work reliably because some compilers only report a warning,
not an error. Amazingly, this has gone undetected for years, even
though it certainly breaks our detection of whether strerror_r
succeeded.
Let's instead test whether the compiler will take the result of
strerror_r() as a switch() argument. It's possible this won't
work universally either, but it's the best idea I could come up with
on the spur of the moment.
Back-patch of commit 751f532b9. Buildfarm results indicate that only
icc-on-Linux actually has an issue here; perhaps the lack of field
reports indicates that people don't build PG for production that way.
Discussion: https://postgr.es/m/10877.1537993279@sss.pgh.pa.us
When the checkpointer receives a SIGHUP signal to update its configuration,
it may need to update the shared memory for full_page_writes and need to
write a WAL record for it. Now, it is quite possible that the XLOG
machinery has not been initialized by that time and it will lead to
assertion failure while doing that. Fix is to allow the initialization of
the XLOG machinery outside critical section.
This bug has been introduced by the commit 2c03216d83 which added the XLOG
machinery initialization in RecoveryInProgress code path.
Reported-by: Dilip Kumar
Author: Dilip Kumar
Reviewed-by: Michael Paquier and Amit Kapila
Backpatch-through: 9.5
Discussion: https://postgr.es/m/CAFiTN-u4BA8KXcQUWDPNgaKAjDXC=C2whnzBM8TAcv=stckYUw@mail.gmail.com
A restart point or a checkpoint recycling WAL segments treats segments
marked with neither ".done" (archiving is done) or ".ready" (segment is
ready to be archived) in archive_status the same way for archive_mode
being "on" or "always". While for a primary this is fine, a standby
running a restart point with archive_mode = on would try to mark such a
segment as ready for archiving, which is something that will never
happen except after the standby is promoted.
Note that this problem applies only to WAL segments coming from the
local pg_wal the first time archive recovery is run. Segments part of a
self-contained base backup are the most common case where this could
happen, however even in this case normally the .done markers would be
most likely part of the backup. Segments recovered from an archive are
marked as .ready or .done by the startup process, and segments finished
streaming are marked as such by the WAL receiver, so they are handled
already.
Reported-by: Haruka Takatsuka
Author: Michael Paquier
Discussion: https://postgr.es/m/15402-a453c90ed4cf88b2@postgresql.org
Backpatch-through: 9.5, where archive_mode = always has been added.
When a table ownership is changed, we must apply that also to any owned
sequences. (Otherwise, it would result in a situation that cannot be
restored, because linked sequences must have the same owner as the
table.) But this was previously only applied to regular tables and
materialized views. But it should also apply to at least foreign
tables. This patch removes the relkind check altogether, because it
doesn't save very much and just introduces the possibility of similar
omissions.
Bug: #15238
Reported-by: Christoph Berg <christoph.berg@credativ.de>
The activation and deactivation of commit timestamp tracking has not
been handled consistently for a primary or standbys at recovery. The
facility can be activated at three different moments of recovery:
- The beginning, where a primary would use the GUC value for the
decision-making, and where a standby relies on the contents of the
control file.
- When replaying a XLOG_PARAMETER_CHANGE record at redo.
- The end, where both primary and standby rely on the GUC value.
Using the GUC value for a primary at the beginning of recovery causes
problems with commit timestamp access when doing crash recovery.
Particularly, when replaying transaction commits, it could be possible
that an attempt to read commit timestamps is done for a transaction
which committed at a moment when track_commit_timestamp was disabled.
A test case is added to reproduce the failure. The test works down to
v11 as it takes advantage of transaction commits within procedures.
Reported-by: Hailong Li
Author: Masahiko Sawasa, Michael Paquier
Reviewed-by: Kyotaro Horiguchi
Discussion: https://postgr.es/m/11224478-a782-203b-1f17-e4797b39bdf0@qunar.com
Backpatch-through: 9.5, where commit timestamps have been introduced.
Apple's latest rearrangements of the system-supplied headers have broken
building of PL/Perl and PL/Tcl. The only practical way to fix PL/Tcl is to
start using the "-isysroot" compiler flag to point to SDK-supplied headers,
as Apple expects. We must also start distinguishing where to find Perl's
headers from where to find its shared library; but that seems like good
cleanup anyway.
Extensions that formerly did something like -I$(perl_archlibexp)/CORE
should now do -I$(perl_includedir)/CORE instead. perl_archlibexp
is still the place to look for libperl.so, though.
If for some reason you don't like the default -isysroot setting, you can
override that by setting PG_SYSROOT in configure's arguments. I don't
currently think people would need to do so, unless maybe for cross-version
build purposes.
In addition, teach configure where to find tclConfig.sh. Our traditional
method of searching $auto_path hasn't worked for the last couple of macOS
releases, and it now seems clear that Apple's not going to change that.
The workaround of manually specifying --with-tclconfig was annoying
already, but Mojave's made it a lot more so because the sysroot path now
has to be included as well. Let's just wire the knowledge into configure
instead. To avoid breaking builds against non-default Tcl installations
(e.g. MacPorts) wherein the $auto_path method probably still works,
arrange to try the additional case only after all else has failed.
Back-patch to all supported versions, since at least the buildfarm
cares about that. The changes are set up to not do anything on macOS
releases that are old enough to not have functional sysroot trees.
array_out overestimated the space needed for its output, possibly by
a very substantial amount if the array is multi-dimensional, because
of wrong order of operations in the loop that counts the number of
curly-brace pairs needed. While the output string is normally
short-lived, this could still cause problems in extreme cases.
An additional minor error was that it counted one more delimiter than
is actually needed.
Repair those errors, add an Assert that the space is now correctly
calculated, and make some minor improvements in the comments.
I also failed to resist the temptation to get rid of an integer
modulus operation per array element; a simple comparison is sufficient.
This bug dates clear back to Berkeley days, so back-patch to all
supported versions.
Keiichi Hirobe, minor additional work by me
Discussion: https://postgr.es/m/CAH=EFxE9W0tRvQkixR2XJRRCToUYUEDkJZk6tnADXugPBRdcdg@mail.gmail.com
This removes a difference between the standard IsUnderPostmaster
execution environment and that of --boot and --single. In a stand-alone
backend, "SELECT random()" always started at the same seed.
On a system capable of using posix shared memory, initdb could still
conclude "selecting dynamic shared memory implementation ... sysv".
Crashed --boot or --single postgres processes orphaned shared memory
objects having names that collided with the not-actually-random names
that initdb probed. The sysv fallback appeared after ten crashes of
--boot or --single postgres. Since --boot and --single are rare in
production use, systems used for PostgreSQL development are the
principal candidate to notice this symptom.
Back-patch to 9.3 (all supported versions). PostgreSQL 9.4 introduced
dynamic shared memory, but 9.3 does share the "SELECT random()" problem.
Reviewed by Tom Lane and Kyotaro HORIGUCHI.
Discussion: https://postgr.es/m/20180915221546.GA3159382@rfd.leadboat.com
In a case where we have multiple relation-scan nodes in a cursor plan,
such as a scan of an inheritance tree, it's possible to fetch from a
given scan node, then rewind the cursor and fetch some row from an
earlier scan node. In such a case, execCurrent.c mistakenly thought
that the later scan node was still active, because ExecReScan hadn't
done anything to make it look not-active. We'd get some sort of
failure in the case of a SeqScan node, because the node's scan tuple
slot would be pointing at a HeapTuple whose t_self gets reset to
invalid by heapam.c. But it seems possible that for other relation
scan node types we'd actually return a valid tuple TID to the caller,
resulting in updating or deleting a tuple that shouldn't have been
considered current. To fix, forcibly clear the ScanTupleSlot in
ExecScanReScan.
Another issue here, which seems only latent at the moment but could
easily become a live bug in future, is that rewinding a cursor does
not necessarily lead to *immediately* applying ExecReScan to every
scan-level node in the plan tree. Upper-level nodes will think that
they can postpone that call if their child node is already marked
with chgParam flags. I don't see a way for that to happen today in
a plan tree that's simple enough for execCurrent.c's search_plan_tree
to understand, but that's one heck of a fragile assumption. So, add
some logic in search_plan_tree to detect chgParam flags being set on
nodes that it descended to/through, and assume that that means we
should consider lower scan nodes to be logically reset even if their
ReScan call hasn't actually happened yet.
Per bug #15395 from Matvey Arye. This has been broken for a long time,
so back-patch to all supported branches.
Discussion: https://postgr.es/m/153764171023.14986.280404050547008575@wrigleys.postgresql.org
Several users of extensions complained of crashes in parallel workers
that turned out to be due to syscache access from their _PG_init()
functions. Reorder the initialization of parallel workers so that
libraries are restored after the caches are initialized, and inside a
transaction.
This was reported in bug #15350 and elsewhere. We don't consider it
to be a bug: extensions shouldn't do that, because then they can't be
used in shared_preload_libraries. However, it's a fairly obscure
hazard and these extensions worked in practice before parallel query
came along. So let's make it work. Later commits might add a warning
message and eventually an error.
Back-patch to 9.6, where parallel query landed.
Author: Thomas Munro
Reviewed-by: Amit Kapila
Reported-by: Kieran McCusker, Jimmy
Discussion: https://postgr.es/m/153512195228.1489.8545997741965926448%40wrigleys.postgresql.org
Commit 37c54863c removed the code in StandbyAcquireAccessExclusiveLock
that checked the return value of LockAcquireExtended. That created a
bug, because it's still passing reportMemoryError = false to
LockAcquireExtended, meaning that LOCKACQUIRE_NOT_AVAIL will be returned
if we're out of shared memory for the lock table.
In such a situation, the startup process would believe it had acquired an
exclusive lock even though it hadn't, with potentially dire consequences.
To fix, just drop the use of reportMemoryError = false, which allows us
to simplify the call into a plain LockAcquire(). It's unclear that the
locktable-full situation arises often enough that it's worth having a
better recovery method than crash-and-restart. (I strongly suspect that
the only reason the code path existed at all was that it was relatively
simple to do in the pre-37c54863c implementation. But now it's not.)
LockAcquireExtended's reportMemoryError parameter is now dead code and
could be removed. I refrained from doing so, however, because there
was some interest in resurrecting the behavior if we do get reports of
locktable-full failures in the field. Also, it seems unwise to remove
the parameter concurrently with shipping commit f868a8143, which added a
parameter; if there are any third-party callers of LockAcquireExtended,
we want them to get a wrong-number-of-parameters compile error rather
than a possibly-silent misinterpretation of its last parameter.
Back-patch to 9.6 where the bug was introduced.
Discussion: https://postgr.es/m/6202.1536359835@sss.pgh.pa.us
Chris Travers reported that the startup process can repeatedly try to
cancel a backend that is in a posix_fallocate()/EINTR loop and cause it
to loop forever. Teach the retry loop to give up if an interrupt is
pending. Don't actually check for interrupts in that loop though,
because a non-local exit would skip some clean-up code in the caller.
Back-patch to 9.4 where DSM was added (and posix_fallocate() was later
back-patched).
Author: Chris Travers
Reviewed-by: Ildar Musin, Murat Kabilov, Oleksii Kliukin
Tested-by: Oleksii Kliukin
Discussion: https://postgr.es/m/CAN-RpxB-oeZve_J3SM_6%3DHXPmvEG%3DHX%2B9V9pi8g2YR7YW0rBBg%40mail.gmail.com
The EvalPlanQual machinery assumes that any initplans (that is,
uncorrelated sub-selects) used during an EPQ recheck would have already
been evaluated during the main query; this is implicit in the fact that
execPlan pointers are not copied into the EPQ estate's es_param_exec_vals.
But it's possible for that assumption to fail, if the initplan is only
reached conditionally. For example, a sub-select inside a CASE expression
could be reached during a recheck when it had not been previously, if the
CASE test depends on a column that was just updated.
This bug is old, appearing to date back to my rewrite of EvalPlanQual in
commit 9f2ee8f28, but was not detected until Kyle Samson reported a case.
To fix, force all not-yet-evaluated initplans used within the EPQ plan
subtree to be evaluated at the start of the recheck, before entering the
EPQ environment. This could be inefficient, if such an initplan is
expensive and goes unused again during the recheck --- but that's piling
one layer of improbability atop another. It doesn't seem worth adding
more complexity to prevent that, at least not in the back branches.
It was convenient to use the new-in-v11 ExecEvalParamExecParams function
to implement this, but I didn't like either its name or the specifics of
its API, so revise that.
Back-patch all the way. Rather than rewrite the patch to avoid depending
on bms_next_member() in the oldest branches, I chose to back-patch that
function into 9.4 and 9.3. (This isn't the first time back-patches have
needed that, and it exhausted my patience.) I also chose to back-patch
some test cases added by commits 71404af2a and 342a1ffa2 into 9.4 and 9.3,
so that the 9.x versions of eval-plan-qual.spec are all the same.
Andrew Gierth diagnosed the problem and contributed the added test cases,
though the actual code changes are by me.
Discussion: https://postgr.es/m/A033A40A-B234-4324-BE37-272279F7B627@tripadvisor.com
Allowing sub-select containing LIMIT/OFFSET in workers can lead to
inconsistent results at the top-level as there is no guarantee that the
row order will be fully deterministic. The fix is to prohibit pushing
LIMIT/OFFSET within sub-selects to workers.
Reported-by: Andrew Fletcher
Bug: 15324
Author: Amit Kapila
Reviewed-by: Dilip Kumar
Backpatch-through: 9.6
Discussion: https://postgr.es/m/153417684333.10284.11356259990921828616@wrigleys.postgresql.org
XLogInsert fails to attach a required FPI to the first record after
full_page_writes is turned on by the last checkpoint. This bug got
introduced in 9.5 due to code rearrangement in commits 2c03216d83 and
2076db2aea. Fix it by ensuring that XLogInsertRecord performs a
recomputation when the given record is generated with FPW as off but
found that the flag has been turned on while actually inserting the
record.
Reported-by: Kyotaro Horiguchi
Author: Kyotaro Horiguchi
Reviewed-by: Amit Kapila
Backpatch-through: 9.5 where this problem was introduced
Discussion: https://postgr.es/m/20180420.151043.74298611.horiguchi.kyotaro@lab.ntt.co.jp
Commit c8ea87e4b introduced a temporary conversion buffer for
substrings extracted during regexp splits. Unfortunately the code that
sized it was failing to ignore the effects of ignored degenerate
regexp matches, so for regexp_split_* calls it could under-size the
buffer in such cases.
Fix, and add some regression test cases (though those will only catch
the bug if run in a multibyte encoding).
Backpatch to 9.3 as the faulty code was.
Thanks to the PostGIS project, Regina Obe and Paul Ramsey for the
report (via IRC) and assistance in analysis. Patch by me.
spgrescan would first reset traversalCxt, and then traverse a
potentially non-empty stack containing pointers to traversalValues
which had been allocated in those contexts, freeing them a second
time. This bug originates in commit ccd6eb49a where traversalValue was
introduced.
Repair by traversing the stack before the context reset; this isn't
ideal, since it means doing retail pfree in a context that's about to
be reset, but the freeing of a stack entry is also done in other
places in the code during the scan so it's not worth trying to
refactor it further. Regression test added.
Backpatch to 9.6 where the problem was introduced.
Per bug #15378; analysis and patch by me, originally from a report on
IRC by user velix; see also PostGIS ticket #4174; review by Alexander
Korotkov.
Discussion: https://postgr.es/m/153663176628.23136.11901365223750051490@wrigleys.postgresql.org
ginRedoRecompress() replays actions over compressed segments of posting list
in-place. However, it might lead to write past pg_upper, because intermediate
state during playing the changes can take more space than both original state
and final state. This commit fixes that by refuse from in-place modification.
Instead page tail is copied once modification is started, and then it's used
as the source of original segments. Backpatch to 9.4 where posting list
compression was introduced.
Reported-by: Sivasubramanian Ramasubramanian
Discussion: https://postgr.es/m/1536091151804.6588%40amazon.com
Author: Alexander Korotkov based on patch from and ideas by Sivasubramanian Ramasubramanian
Review: Sivasubramanian Ramasubramanian
Backpatch-through: 9.4
This patch removes two sources of interference between nominally
independent functions when one SPI-using function calls another,
perhaps without knowing that it does so.
Chapman Flack pointed out that xml.c's query_to_xml_internal() expects
SPI_tuptable and SPI_processed to stay valid across datatype output
function calls; but it's possible that such a call could involve
re-entrant use of SPI. It seems likely that there are similar hazards
elsewhere, if not in the core code then in third-party SPI users.
Previously SPI_finish() reset SPI's API globals to zeroes/nulls, which
would typically make for a crash in such a situation. Restoring them
to the values they had at SPI_connect() seems like a considerably more
useful behavior, and it still meets the design goal of not leaving any
dangling pointers to tuple tables of the function being exited.
Also, cause SPI_connect() to reset these variables to zeroes/nulls after
saving them. This prevents interference in the opposite direction: it's
possible that a SPI-using function that's only ever been tested standalone
contains assumptions that these variables start out as zeroes. That was
the case as long as you were the outermost SPI user, but not so much for
an inner user. Now it's consistent.
Report and fix suggestion by Chapman Flack, actual patch by me.
Back-patch to all supported branches.
Discussion: https://postgr.es/m/9fa25bef-2e4f-1c32-22a4-3ad0723c4a17@anastigmatix.net
It's somewhat surprising that we got away with this before. (Actually,
since nobody tests this routinely AFAIK, it might've been broken for
awhile. But it's definitely broken in the wake of commit f868a8143.)
It seems sufficient to limit the forced recursion to a small number
of levels.
Back-patch to all supported branches, like the preceding patch.
Discussion: https://postgr.es/m/12259.1532117714@sss.pgh.pa.us
LockRelationOid and sibling routines supposed that, if our session already
holds the lock they were asked to acquire, they could skip calling
AcceptInvalidationMessages on the grounds that we must have already read
any remote sinval messages issued against the relation being locked.
This is normally true, but there's a critical special case where it's not:
processing inside AcceptInvalidationMessages might attempt to access system
relations, resulting in a recursive call to acquire a relation lock.
Hence, if the outer call had acquired that same system catalog lock, we'd
fall through, despite the possibility that there's an as-yet-unread sinval
message for that system catalog. This could, for example, result in
failure to access a system catalog or index that had just been processed
by VACUUM FULL. This is the explanation for buildfarm failures we've been
seeing intermittently for the past three months. The bug is far older
than that, but commits a54e1f158 et al added a new recursion case within
AcceptInvalidationMessages that is apparently easier to hit than any
previous case.
To fix this, we must not skip calling AcceptInvalidationMessages until
we have *finished* a call to it since acquiring a relation lock, not
merely acquired the lock. (There's already adequate logic inside
AcceptInvalidationMessages to deal with being called recursively.)
Fortunately, we can implement that at trivial cost, by adding a flag
to LOCALLOCK hashtable entries that tracks whether we know we have
completed such a call.
There is an API hazard added by this patch for external callers of
LockAcquire: if anything is testing for LOCKACQUIRE_ALREADY_HELD,
it might be fooled by the new return code LOCKACQUIRE_ALREADY_CLEAR
into thinking the lock wasn't already held. This should be a fail-soft
condition, though, unless something very bizarre is being done in
response to the test.
Also, I added an additional output argument to LockAcquireExtended,
assuming that that probably isn't called by any outside code given
the very limited usefulness of its additional functionality.
Back-patch to all supported branches.
Discussion: https://postgr.es/m/12259.1532117714@sss.pgh.pa.us
workers.
Allowing window function calculation in workers leads to inconsistent
results because if the input row ordering is not fully deterministic, the
output of window functions might vary across workers. The fix is to treat
them as parallel-restricted.
In the passing, improve the coding pattern in max_parallel_hazard_walker
so that it has a chain of mutually-exclusive if ... else if ... else if
... else if ... IsA tests.
Reported-by: Marko Tiikkaja
Bug: 15324
Author: Amit Kapila
Reviewed-by: Tom Lane
Backpatch-through: 9.6
Discussion: https://postgr.es/m/CAL9smLAnfPJCDUUG4ckX2iznj53V7VSMsYefzZieN93YxTNOcw@mail.gmail.com
At the beginning of recovery, information from replication slots is
recovered from disk to memory. In order to ensure the durability of the
information, the status file as well as its parent directory are
synced. It happens that the sync on the parent directory was done
directly using the status file path, which is logically incorrect, and
the current code has been doing a sync on the same object twice in a
row.
Reported-by: Konstantin Knizhnik
Diagnosed-by: Konstantin Knizhnik
Author: Michael Paquier
Discussion: https://postgr.es/m/9eb1a6d5-b66f-2640-598d-c5ea46b8f68a@postgrespro.ru
Backpatch-through: 9.4-
There's a project policy against using plain "char buf[BLCKSZ]" local
or static variables as page buffers; preferred style is to palloc or
malloc each buffer to ensure it is MAXALIGN'd. However, that policy's
been ignored in an increasing number of places. We've apparently got
away with it so far, probably because (a) relatively few people use
platforms on which misalignment causes core dumps and/or (b) the
variables chance to be sufficiently aligned anyway. But this is not
something to rely on. Moreover, even if we don't get a core dump,
we might be paying a lot of cycles for misaligned accesses.
To fix, invent new union types PGAlignedBlock and PGAlignedXLogBlock
that the compiler must allocate with sufficient alignment, and use
those in place of plain char arrays.
I used these types even for variables where there's no risk of a
misaligned access, since ensuring proper alignment should make
kernel data transfers faster. I also changed some places where
we had been palloc'ing short-lived buffers, for coding style
uniformity and to save palloc/pfree overhead.
Since this seems to be a live portability hazard (despite the lack
of field reports), back-patch to all supported versions.
Patch by me; thanks to Michael Paquier for review.
Discussion: https://postgr.es/m/1535618100.1286.3.camel@credativ.de
Healthy clients of servers having poor I/O performance, such as
buildfarm members hamster and tern, saw unexpected timeouts. That
disagreed with documentation. This fix adds one gettimeofday() call
whenever ProcessRepliesIfAny() finds no client reply messages.
Back-patch to 9.4; the bug's symptom is rare and mild, and the code all
moved between 9.3 and 9.4.
Discussion: https://postgr.es/m/20180826034600.GA1105084@rfd.leadboat.com
Startup process has improved its calculation of incorrect minimum
consistent point in 8d68ee6, which ensures that all WAL available gets
replayed when doing crash recovery, and has introduced an incorrect
calculation of the minimum recovery point for non-startup processes,
which can cause incorrect page references on a standby when for example
the background writer flushed a couple of pages on-disk but was not
updating the control file to let a subsequent crash recovery replay to
where it should have.
The only case where this has been reported to be a problem is when a
standby needs to calculate the latest removed xid when replaying a btree
deletion record, so one would need connections on a standby that happen
just after recovery has thought it reached a consistent point. Using a
background worker which is started after the consistent point is reached
would be the easiest way to get into problems if it connects to a
database. Having clients which attempt to connect periodically could
also be a problem, but the odds of seeing this problem are much lower.
The fix used is pretty simple, as the idea is to give access to the
minimum recovery point written in the control file to non-startup
processes so as they use a reference, while the startup process still
initializes its own references of the minimum consistent point so as the
original problem with incorrect page references happening post-promotion
with a crash do not show up.
Reported-by: Alexander Kukushkin
Diagnosed-by: Alexander Kukushkin
Author: Michael Paquier
Reviewed-by: Kyotaro Horiguchi, Alexander Kukushkin
Discussion: https://postgr.es/m/153492341830.1368.3936905691758473953@wrigleys.postgresql.org
Backpatch-through: 9.3
In general, Postgres requires -fno-strict-aliasing with compilers that
implement C99 strict aliasing rules. There's little hope of getting
rid of that overall. But it seems like it would be a good idea if
storage/checksum_impl.h in particular didn't depend on it, because
that header is explicitly intended to be included by external programs.
We don't have a lot of control over the compiler switches that an
external program might use, as shown by Michael Banck's report of
failure in a privately-modified version of pg_verify_checksums.
Hence, switch to using a union in place of willy-nilly pointer casting
inside this file. I think this makes the code a bit more readable
anyway.
checksum_impl.h hasn't changed since it was introduced in 9.3,
so back-patch all the way.
Discussion: https://postgr.es/m/1535618100.1286.3.camel@credativ.de
When a postmaster gets into its phase PM_STARTUP, it would start
background workers using BgWorkerStart_PostmasterStart mode immediately,
which would cause problems for a fast shutdown as the postmaster forgets
to send SIGTERM to already-started background workers. With smart and
immediate shutdowns, this correctly happened, and fast shutdown is the
only mode missing the shot.
Author: Alexander Kukushkin
Reviewed-by: Michael Paquier
Discussion: https://postgr.es/m/CAFh8B=mvnD8+DZUfzpi50DoaDfZRDfd7S=gwj5vU9GYn8UvHkA@mail.gmail.com
Backpatch-through: 9.5
This function had a blacklist of dump object types that it believed
needed exclusive lock ... but we hadn't maintained that, so that it
was missing ROW SECURITY, POLICY, and INDEX ATTACH items, all of
which need (or should be treated as needing) exclusive lock.
Since the same oversight seems likely in future, let's reverse the
sense of the test so that the code has a whitelist of safe object
types; better to wrongly assume a command can't be run in parallel
than the opposite. Currently the only POST_DATA object type that's
safe is CREATE INDEX ... and that list hasn't changed in a long time.
Back-patch to 9.5 where RLS came in.
Discussion: https://postgr.es/m/11450.1535483506@sss.pgh.pa.us
regexp_matches, regexp_split_to_table and regexp_split_to_array all
work by compiling a list of match positions as character offsets (NOT
byte positions) in the source string.
Formerly, they then used text_substr to extract the matched text; but
in a multi-byte encoding, that counts the characters in the string,
and the characters needed to reach the starting byte position, on
every call. Accordingly, the performance degraded as the product of
the input string length and the number of match positions, such that
splitting a string of a few hundred kbytes could take many minutes.
Repair by keeping the wide-character copy of the input string
available (only in the case where encoding_max_length is not 1) after
performing the match operation, and extracting substrings from that
instead. This reduces the complexity to being linear in the number of
result bytes, discounting the actual regexp match itself (which is not
affected by this patch).
In passing, remove cleanup using retail pfree() which was obsoleted by
commit ff428cded (Feb 2008) which made cleanup of SRF multi-call
contexts automatic. Also increase (to ~134 million) the maximum number
of matches and provide an error message when it is reached.
Backpatch all the way because this has been wrong forever.
Analysis and patch by me; review by Kaiting Chen.
Discussion: https://postgr.es/m/87pnyn55qh.fsf@news-spur.riddles.org.uk
see also https://postgr.es/m/87lg996g4r.fsf@news-spur.riddles.org.uk
The archive should show a dependency on the item's table, but it failed
to include one. This could cause failures in parallel restore due to
emitting ALTER TABLE ... ENABLE ROW LEVEL SECURITY before restoring
the table's data. In practice the odds of a problem seem low, since
you would typically need to have set FORCE ROW LEVEL SECURITY as well,
and you'd also need a very high --jobs count to have any chance of this
happening. That probably explains the lack of field reports.
Still, it's a bug, so back-patch to 9.5 where RLS was introduced.
Discussion: https://postgr.es/m/19784.1535390902@sss.pgh.pa.us
The previous coding figured it'd be good enough to postpone opening
the first CSV log file until we got a message we needed to write there.
This is unsafe, though, because if the open fails we end up in infinite
recursion trying to report the failure. Instead make the CSV log file
management code look as nearly as possible like the longstanding logic
for the stderr log file. In particular, open it immediately at postmaster
startup (if enabled), or when we get a SIGHUP in which we find that
log_destination has been changed to enable CSV logging.
It seems OK to fail if a postmaster-start-time open attempt fails, as
we've long done for the stderr log file. But we can't die if we fail
to open a CSV log file during SIGHUP, so we're still left with a problem.
In that case, write any output meant for the CSV log file to the stderr
log file. (This will also cover race-condition cases in which backends
send CSV log data before or after we have the CSV log file open.)
This patch also fixes an ancient oversight that, if CSV logging was
turned off during a SIGHUP, we never actually closed the last CSV
log file.
In passing, remember to reset whereToSendOutput = DestNone during syslogger
start, since (unlike all other postmaster children) it's forked before the
postmaster has done that. This made for a platform-dependent difference
in error reporting behavior between the syslogger and other children:
except on Windows, it'd report problems to the original postmaster stderr
as well as the normal error log file(s). It's barely possible that that
was intentional at some point; but it doesn't seem likely to be desirable
in production, and the platform dependency definitely isn't desirable.
Per report from Alexander Kukushkin. It's been like this for a long time,
so back-patch to all supported branches.
Discussion: https://postgr.es/m/CAFh8B==iLUD_gqC-dAENS0V+kVrCeGiKujtKqSQ7++S-caaChw@mail.gmail.com
Commits c6b3c939b (which fixed the precedence of >=, <=, <> operators)
and 865f14a2d (which added support for the standard => notation for
named arguments) created a class of lexer tokens which look like
multi-character operators but which have their own token IDs distinct
from Op. However, longest-match rules meant that following any of
these tokens with another operator character, as in (1<>-1), would
cause them to be incorrectly returned as Op.
The error here isn't immediately obvious, because the parser would
usually still find the correct operator via the Op token, but there
were more subtle problems:
1. If immediately followed by a comment or +-, >= <= <> would be given
the old precedence of Op rather than the correct new precedence;
2. If followed by a comment, != would be returned as Op rather than as
NOT_EQUAL, causing it not to be found at all;
3. If followed by a comment or +-, the => token for named arguments
would be lexed as Op, causing the argument to be mis-parsed as a
simple expression, usually causing an error.
Fix by explicitly checking for the operators in the {operator} code
block in addition to all the existing special cases there.
Backpatch to 9.5 where the problem was introduced.
Analysis and patch by me; review by Tom Lane.
Discussion: https://postgr.es/m/87va851ppl.fsf@news-spur.riddles.org.uk
The lexer's handling of operators contained an O(N^3) hazard when
dealing with long strings of + or - characters; it seems hard to
prevent this case from being O(N^2), but the additional N multiplier
was not needed.
Backpatch all the way since this has been there since 7.x, and it
presents at least a mild hazard in that trying to do Bind, PREPARE or
EXPLAIN on a hostile query could take excessive time (without
honouring cancels or timeouts) even if the query was never executed.
While monitoring the code, a couple of issues related to string
translation has showed up:
- Some routines for auto-updatable views return an error string, which
sometimes missed the shot. A comment regarding string translation is
added for each routine to help with future features.
- GSSAPI authentication missed two translations.
- vacuumdb handles non-translated strings.
Reported-by: Kyotaro Horiguchi
Author: Kyotaro Horiguchi
Reviewed-by: Michael Paquier, Tom Lane
Discussion: https://postgr.es/m/20180810.152131.31921918.horiguchi.kyotaro@lab.ntt.co.jp
Backpatch-through: 9.3
Previously, this code blindly followed the common coding pattern of
passing PQserverVersion(AH->connection) as the server-version parameter
of fmtQualifiedId. That works as long as we have a connection; but in
pg_restore with text output, we don't. Instead we got a zero from
PQserverVersion, which fmtQualifiedId interpreted as "server is too old to
have schemas", and so the name went unqualified. That still accidentally
managed to work in many cases, which is probably why this ancient bug went
undetected for so long. It only became obvious in the wake of the changes
to force dump/restore to execute with restricted search_path.
In HEAD/v11, let's deal with this by ripping out fmtQualifiedId's server-
version behavioral dependency, and just making it schema-qualify all the
time. We no longer support pg_dump from servers old enough to need the
ability to omit schema name, let alone restoring to them. (Also, the few
callers outside pg_dump already didn't work with pre-schema servers.)
In older branches, that's not an acceptable solution, so instead just
tweak the DISABLE/ENABLE TRIGGER logic to ensure it will schema-qualify
its output regardless of server version.
Per bug #15338 from Oleg somebody. Back-patch to all supported branches.
Discussion: https://postgr.es/m/153452458706.1316.5328079417086507743@wrigleys.postgresql.org
When executing a SubPlan in an expression, the EState's direction
field was left alone, resulting in an attempt to execute the subplan
backwards if it was encountered during a backwards scan of a cursor.
Also, though much less likely, it was possible to reach the execution
of an InitPlan while in backwards-scan state.
Repair by saving/restoring estate->es_direction and forcing forward
scan mode in the relevant places.
Backpatch all the way, since this has been broken since 8.3 (prior to
commit c7ff7663e, SubPlans had their own EStates rather than sharing
the parent plan's, so there was no confusion over scan direction).
Per bug #15336 reported by Vladimir Baranoff; analysis and patch by
me, review by Tom Lane.
Discussion: https://postgr.es/m/153449812167.1304.1741624125628126322@wrigleys.postgresql.org
The function was forgetting to close the file descriptor, resulting
in failures like this:
ERROR: 53000: exceeded maxAllocatedDescs (492) while trying to open
file "pg_logical/mappings/map-4000-4eb-1_60DE1E08-5376b5-537c6b"
LOCATION: OpenTransientFile, fd.c:2161
Simply close the file at the end, and backpatch to 9.4 (where logical
decoding was introduced). While at it, fix a nearby typo.
Discussion: https://www.postgresql.org/message-id/flat/738a590a-2ce5-9394-2bef-7b1caad89b37%402ndquadrant.com
C99 says that the result should be the number of bytes that would have
been emitted given a large enough buffer, not the number we actually
were able to put in the buffer. It's time to make our substitute
implementation comply with that. Not doing so results in inefficiency
in buffer-enlargement cases, and also poses a portability hazard for
third-party code that might expect C99-compliant snprintf behavior
within Postgres.
In passing, remove useless tests for str == NULL; neither C99 nor
predecessor standards ever allowed that except when count == 0,
so I see no reason to expend cycles on making that a non-crash case
for this implementation. Also, don't waste a byte in pg_vfprintf's
local I/O buffer; this might have performance benefits by allowing
aligned writes during flushbuffer calls.
Back-patch of commit 805889d7d. There was some concern about this
possibly breaking code that assumes pre-C99 behavior, but there is
much more risk (and reality, in our own code) of code that assumes
C99 behavior and hence fails to detect buffer overrun without this.
Discussion: https://postgr.es/m/17245.1534289329@sss.pgh.pa.us
We aren't very strict about keeping FSM up to date on WAL replay,
because per-page freespace values aren't critical in replicas (can't
write to heap in a replica; and if the replica is promoted, the values
would be updated by VACUUM anyway). However, VACUUM since 9.6 can skip
processing pages marked all-visible or all-frozen, and if such pages are
recorded in FSM with wrong values, those values are blindly propagated
to FSM's upper layers by VACUUM's FreeSpaceMapVacuum. (This rationale
assumes that crashes are not very frequent, because those would cause
outdated FSM to occur in the primary.)
Even when the FSM is outdated in standby, things are not too bad
normally, because, most per-page FSM values will be zero (other than
those propagated with the base-backup that created the standby); only
once the remaining free space is less than 0.2*BLCKSZ the per-page value
is maintained by WAL replay of heap ins/upd/del. However, if
wal_log_hints=on causes complete FSM pages to be propagated to a standby
via full-page images, many too-optimistic per-page values can end up
being registered in the standby.
Incorrect per-page values aren't critical in most cases, since an
inserter that is given a page that doesn't actually contain the claimed
free space will update FSM with the correct value, and retry until it
finds a usable page. However, if there are many such updates to do, an
inserter can spend a long time doing them before a usable page is found;
in a heavily trafficked insert-only table with many concurrent inserters
this has been observed to cause several second stalls, causing visible
application malfunction.
To fix this problem, it seems sufficient to have heap_xlog_visible
(replay of setting all-visible and all-frozen VM bits for a heap page)
update the FSM value for the page being processed. This fixes the
per-page counters together with making the page skippable to vacuum, so
when vacuum does FreeSpaceMapVacuum, the values propagated to FSM upper
layers are the correct ones, avoiding the problem.
While at it, apply the same fix to heap_xlog_clean (replay of tuple
removal by HOT pruning and vacuum). This makes any space freed by the
cleaning available earlier than the next vacuum in the promoted replica.
Backpatch to 9.6, where this problem was diagnosed on an insert-only
table with all-frozen pages, which were introduced as a concept in that
release. Theoretically it could apply with all-visible pages to older
branches, but there's been no report of that and it doesn't backpatch
cleanly anyway.
Author: Álvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/20180802172857.5skoexsilnjvgruk@alvherre.pgsql
Fix a small number of places that were testing the result of snprintf()
but doing so incorrectly. The right test for buffer overrun, per C99,
is "result >= bufsize" not "result > bufsize". Some places were also
checking for failure with "result == -1", but the standard only says
that a negative value is delivered on failure.
(Note that this only makes these places correct if snprintf() delivers
C99-compliant results. But at least now these places are consistent
with all the other places where we assume that.)
Also, make psql_start_test() and isolation_start_test() check for
buffer overrun while constructing their shell commands. There seems
like a higher risk of overrun, with more severe consequences, here
than there is for the individual file paths that are made elsewhere
in the same functions, so this seemed like a worthwhile change.
Also fix guc.c's do_serialize() to initialize errno = 0 before
calling vsnprintf. In principle, this should be unnecessary because
vsnprintf should have set errno if it returns a failure indication ...
but the other two places this coding pattern is cribbed from don't
assume that, so let's be consistent.
These errors are all very old, so back-patch as appropriate. I think
that only the shell command overrun cases are even theoretically
reachable in practice, but there's not much point in erroneous error
checks.
Discussion: https://postgr.es/m/17245.1534289329@sss.pgh.pa.us
Commit 244142d32a only tested for the
pg_controldata output for primary servers, but standby servers have
different "Database cluster state" output, so check for that too.
Diagnosed-by: Michael Paquier
Discussion: https://postgr.es/m/20180810164240.GM13638@paquier.xyz
Backpatch-through: 9.3