A crash in the middle of a checkpoint with some two-phase state data
already flushed to disk by this checkpoint could cause a follow-up crash
recovery to recover twice the same transaction, once from what has been
found in pg_twophase/ at the beginning of recovery and a second time
when replaying its corresponding record.
This would lead to FATAL failures in the startup process during
recovery, where the same transaction would have a state recovered twice
instead of once:
LOG: recovering prepared transaction 731 from shared memory
LOG: recovering prepared transaction 731 from shared memory
FATAL: lock ExclusiveLock on object 731/0/0 is already held
This issue is fixed by skipping the addition of any 2PC state coming
from a record whose equivalent 2PC state file has already been loaded in
TwoPhaseState at the beginning of recovery by restoreTwoPhaseData(),
which is OK as long as the system has not reached a consistent state.
The timing to get a messed up recovery processing is very racy, and
would very unlikely happen. The thread that has reported the issue has
demonstrated the bug using injection points to force a PANIC in the
middle of a checkpoint.
Issue introduced in 728bd99, so backpatch all the way down.
Reported-by: "suyu.cmj" <mengjuan.cmj@alibaba-inc.com>
Author: "suyu.cmj" <mengjuan.cmj@alibaba-inc.com>
Author: Michael Paquier
Discussion: https://postgr.es/m/109e6994-b971-48cb-84f6-829646f18b4c.mengjuan.cmj@alibaba-inc.com
Backpatch-through: 11
RelationReloadIndexInfo() is a fast-path used for index reloads in the
relation cache, and it has always forgotten about updating
indisreplident, which is something that would happen after an index is
selected for a replica identity. This can lead to incorrect cache
information provided when executing a command in a transaction context
that updates indisreplident.
None of the code paths currently on HEAD that need to check upon
pg_index.indisreplident fetch its value from the relation cache, always
relying on a fresh copy on the syscache. Unfortunately, this may not be
the case of out-of-core code, that could see out-of-date value.
Author: Shruthi Gowda
Reviewed-by: Robert Haas, Dilip Kumar, Michael Paquier
Discussion: https://postgr.es/m/CAASxf_PBcxax0wW-3gErUyftZ0XrCs3Lrpuhq4-Z3Fak1DoW7Q@mail.gmail.com
Backpatch-through: 11
indisvalid is switched to true for partitioned indexes when all its
partitions have valid indexes when attaching a new partition, up to the
top-most parent if all its leaves are themselves valid when dealing with
multiple layers of partitions.
The copy of the tuple from pg_index used to switch indisvalid to true
came from the relation cache, which is incorrect. Particularly, in the
case reported by Shruthi Gowda, executing a series of commands in a
single transaction would cause the validation of partitioned indexes to
use an incorrect version of a pg_index tuple, as indexes are reloaded
after an invalidation request with RelationReloadIndexInfo(), a much
faster version than a full index cache rebuild. In this case, the
limited information updated in the cache leads to an incorrect version
of the tuple used. One of the symptoms reported was the following
error, with a replica identity update, for instance:
"ERROR: attempted to update invisible tuple"
This is incorrect since 8b08f7d, so backpatch all the way down.
Reported-by: Shruthi Gowda
Author: Michael Paquier
Reviewed-by: Shruthi Gowda, Dilip Kumar
Discussion: https://postgr.es/m/CAASxf_PBcxax0wW-3gErUyftZ0XrCs3Lrpuhq4-Z3Fak1DoW7Q@mail.gmail.com
Backpatch-through: 11
Until now, when DROP DATABASE got interrupted in the wrong moment, the removal
of the pg_database row would also roll back, even though some irreversible
steps have already been taken. E.g. DropDatabaseBuffers() might have thrown
out dirty buffers, or files could have been unlinked. But we continued to
allow connections to such a corrupted database.
To fix this, mark databases invalid with an in-place update, just before
starting to perform irreversible steps. As we can't add a new column in the
back branches, we use pg_database.datconnlimit = -2 for this purpose.
An invalid database cannot be connected to anymore, but can still be
dropped.
Unfortunately we can't easily add output to psql's \l to indicate that some
database is invalid, it doesn't fit in any of the existing columns.
Add tests verifying that a interrupted DROP DATABASE is handled correctly in
the backend and in various tools.
Reported-by: Evgeny Morozov <postgresql3@realityexists.net>
Author: Andres Freund <andres@anarazel.de>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/20230509004637.cgvmfwrbht7xm7p6@awork3.anarazel.de
Discussion: https://postgr.es/m/20230314174521.74jl6ffqsee5mtug@awork3.anarazel.de
Backpatch: 11-, bug present in all supported versions
When vac_truncate_clog() encounters bogus datfrozenxid / datminmxid values, it
returns early. Unfortunately, until now, it did not release
WrapLimitsVacuumLock. If the backend later tries to acquire
WrapLimitsVacuumLock, the session / autovacuum worker hangs in an
uncancellable way. Similarly, other sessions will hang waiting for the
lock. However, if the backend holding the lock exited or errored out for some
reason, the lock was released.
The bug was introduced as a side effect of 566372b3d6.
It is interesting that there are no production reports of this problem. That
is likely due to a mix of bugs leading to bogus values having gotten less
common, process exit releasing locks and instances of hangs being hard to
debug for "normal" users.
Discussion: https://postgr.es/m/20230621221208.vhsqgduwfpzwxnpg@awork3.anarazel.de
GiST compress functions (like all GiST opclass functions) are
supposed to be called in short-lived memory contexts, so that
minor memory leaks in them are not of concern, and indeed
explicit pfree's are likely slightly counterproductive.
But this one in g_intbig_compress() is more than
slightly counterproductive, because it's guarded by
"if (in != DatumGetArrayTypeP(entry->key))" which means
that if this test succeeds, we've detoasted the datum twice.
(And to add insult to injury, the extra detoast result is
leaked.) Let's just drop the whole stanza, relying on the
GiST temporary context mechanism to clean up in good time.
The analogous bit in g_int_compress() is
if (r != (ArrayType *) DatumGetPointer(entry->key))
pfree(r);
which doesn't have the gratuitous-detoast problem so
I left it alone. Perhaps there is a case for removing
unnecessary pfree's more widely, but I'm not sure if it's
worth the code churn.
The potential extra decompress seems expensive enough to
justify calling this a (minor) performance bug and
back-patching.
Konstantin Knizhnik, Matthias van de Meent, Tom Lane
Discussion: https://postgr.es/m/CAEze2Wi86=DxErfvf+SCB2UKmU2amKOF60BKuJOX=w-RojRn0A@mail.gmail.com
Since PostgresMain calls sigsetjmp, any local variables that are not
marked "volatile" have a risk of unspecified behavior. In practice
this means that when control returns via longjmp, such variables might
get reset to their values as of the time of sigsetjmp, depending on
whether the compiler chose to put them in registers or on the stack.
We were careful about this for "send_ready_for_query", but not the
other local variables.
In the case of the timeout_enabled flags, resetting them to
their initial "false" states is actually good, since we do
"disable_all_timeouts()" in the longjmp cleanup code path. If that
does not happen, we risk uselessly calling "disable_timeout()" later,
which is harmless but a little bit expensive. Let's explicitly reset
these flags so that the behavior is correct and platform-independent.
(This change means that we really don't need the new "volatile"
markings after all, but let's install them anyway since any change
in this logic could re-introduce a problem.)
There is no issue for "firstchar" and "input_message" because those
are explicitly reinitialized each time through the query processing
loop. To make that clearer, move them to be declared inside the loop.
That leaves us with all the function-lifespan locals except the
sigjmp_buf itself marked as volatile, which seems like a good policy
to have going forward.
Because of the possibility of extra disable_timeout() calls, this
seems worth back-patching.
Sergey Shinderuk and Tom Lane
Discussion: https://postgr.es/m/2eda015b-7dff-47fd-d5e2-f1a9899b90a6@postgrespro.ru
As coded, the code would use as a base comparison the namespace OID from
the first object scanned in pg_depend when switching its namespace
dependency entry to the new one, and use it as a base of comparison for
any follow-up checks. It would also be used as the old namespace OID to
switch *from* for the extension's pg_depend entry. Hence, if the first
object scanned has a namespace different than the one stored in the
extension, we would finish by:
- Not checking that the extension objects map with the extension's
schema.
- Not switching the extension -> namespace dependency entry to the new
namespace provided by the user, making ALTER EXTENSION ineffective.
This issue exists since this command has been introduced in d9572c4 for
relocatable extension, so backpatch all the way down to 11. The test
case has been provided by Heikki, that I have tweaked a bit to show the
effects on pg_depend for the extension.
Reported-by: Heikki Linnakangas
Author: Michael Paquier, Heikki Linnakangas
Discussion: https://postgr.es/m/20eea594-a05b-4c31-491b-007b6fceef28@iki.fi
Backpatch-through: 11
Commit e213de8e78 fixed a problem with path lengths to a tempdir on
Windows, but caused problems on at least some Unix systems where the
system tempdir is on a different file system. To work around this, only
used the system temdir for the destination of pg_replslot on Windows,
and otherwise restore the old behaviour.
Backpatch to relase 14 like the previous patch.
Problem exposed by a myriad of buildfarm animals.
The symlink to a longer location tripped up some Windows limit on
buildfarm animal fairywren when running with meson, which uses slightly
longer paths.
Backpatch to release 14 to keep the script in sync. Before that the
script skipped all symlink related tests on Windows.
On Windows, it's sometimes difficult to create a file with a path longer
than 255 chars, and if it can be created it might not be seen by the
archiver. This can be triggered by the test for tar backups with
filenames greater than 100 bytes. So we skip that test if the path would
exceed 255.
Backpatch to all live branches.
Reviewed by Daniel Gustafsson
Discussion: https://postgr.es/m/666ac55b-3400-fb2c-2cea-0281bf36a53c@dunslane.net
We create a file, so we better WAL-log it. In practice, all the
built-in index AMs and all extensions that I'm aware of write a
metapage to the init fork, which is WAL-logged, and replay of the
metapage implicitly creates the fork too. But if ambuildempty() didn't
write any page, we would miss it.
This can be seen with dummy_index_am. Set up replication, create a
'dummy_index_am' index on an unlogged table, and look at the files
created in the replica: the init fork is not created on the
replica. Dummy_index_am doesn't do anything with the relation files,
however, so it doesn't lead to any user-visible errors.
Backpatch to all supported versions.
Reviewed-by: Robert Haas
Discussion: https://www.postgresql.org/message-id/6e5bbc08-cdfc-b2b3-9e23-1a914b9850a9%40iki.fi
This commit reverts the work done by commits 3ba59ccc89 and 72e78d831a.
Those commits were incorrect in asserting that we never acquire any other
heavy-weight lock after acquring page lock other than relation extension
lock. We can acquire a lock on catalogs while doing catalog look up after
acquring page lock.
This won't impact any existing feature but we need to think some other way
to achieve this before parallelizing other write operations or even
improving the parallelism in vacuum (like allowing multiple workers
for an index).
Reported-by: Jaime Casanova
Author: Amit Kapila
Backpatch-through: 13
Discussion: https://postgr.es/m/CAJKUy5jffnRKNvRHKQ0LynRb0RJC-o4P8Ku3x9vGAVLwDBWumQ@mail.gmail.com
llvm_release_context() called llvm_enter_fatal_on_oom(), but was missing
the corresponding llvm_leave_fatal_on_oom() call. As a result, if JIT was
used at all, we were almost always in the "fatal-on-oom" state.
It only makes a difference if you use an extension written in C++, and
run out of memory in a C++ 'new' call. In that case, you would get a
PostgreSQL FATAL error, instead of the default behavior of throwing a
C++ exception.
Back-patch to all supported versions.
Reviewed-by: Daniel Gustafsson
Discussion: https://www.postgresql.org/message-id/54b78cca-bc84-dad8-4a7e-5b56f764fab5@iki.fi
If you create a table and don't insert any data into it, the relation file
is never fsync'd. You don't lose data, because an empty table doesn't have
any data to begin with, but if you crash and lose the file, subsequent
operations on the table will fail with "could not open file" error.
To fix, register an fsync request in mdcreate(), like we do for mdwrite().
Per discussion, we probably should also fsync the containing directory
after creating a new file. But that's a separate and much wider issue.
Backpatch to all supported versions.
Reviewed-by: Andres Freund, Thomas Munro
Discussion: https://www.postgresql.org/message-id/d47d8122-415e-425c-d0a2-e0160829702d%40iki.fi
The Homebrew package manager changed its default installation prefix
for the new architecture, so a couple of tests need tweaks to find
binaries.
This is a partial backpatch of dc513bc654.
It's OK to be lazy about re-binning memory segments when allocating,
because that can only leave segments in a bin that's too high. We'll
search higher bins if necessary while allocating next time, and
also eventually re-bin, so no memory can become unreachable that way.
However, when freeing memory, the largest contiguous range of free pages
might go up, so we should re-bin eagerly to make sure we don't leave the
segment in a bin that is too low for get_best_segment() to find.
The re-binning code is moved into a function of its own, so it can be
called whenever free pages are returned to the segment's free page map.
Back-patch to all supported releases.
Author: Dongming Liu <ldming101@gmail.com>
Reviewed-by: Robert Haas <robertmhaas@gmail.com> (earlier version)
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/CAL1p7e8LzB2LSeAXo2pXCW4%2BRya9s0sJ3G_ReKOU%3DAjSUWjHWQ%40mail.gmail.com
The ginfast.c code previously checked for conflicts in before locking
the relevant buffer, leaving a window where a RW conflict could be
missed. Re-order.
There was also a place where buffer ID and block number were confused
while trying to predicate-lock a page, noted by visual inspection.
Back-patch to all supported releases. Fixes one more problem discovered
with the reproducer from bug #17949, in this case when Dmitry tried
other index types.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reported-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
When performing a bitmap heap scan, we don't want to miss concurrent
writes that occurred after we observed the heap's rs_nblocks, but before
we took predicate locks on index pages. Therefore, we can't skip
fetching any heap tuples that are referenced by the index, because we
need to test them all with CheckForSerializableConflictOut(). The
old optimization that would ignore any references to blocks >=
rs_nblocks gets in the way of that requirement, because it means that
concurrent writes in that window are ignored.
Removing that optimization shouldn't affect correctness at any isolation
level, because any new tuples shouldn't be visible to an MVCC snapshot.
There also shouldn't be any error-causing references to heap blocks past
the end, because we should have held at least an AccessShareLock on the
table before the index scan. It can't get smaller while our transaction
is running. For now, though, we'll keep the optimization at lower
levels to avoid making unnecessary changes in a bug fix.
Back-patch to all supported releases. In release 11, the code is in a
different place but not fundamentally different. Fixes one aspect of
bug #17949.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
When predicate-locking btrees, we have a special case for completely
empty btrees, since there is no page to lock. This was racy, because,
without buffer lock held, a matching key could be inserted between the
_bt_search() and the PredicateLockRelation() calls.
Fix, by rechecking _bt_search() after taking the relation-level SIREAD
lock, if using SERIALIZABLE isolation and an empty btree is discovered.
Back-patch to all supported releases. Fixes one aspect of bug #17949.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
The test inserted 70k rows into a foreign table, in order to verify
correct behavior with more than 65535 parameters, and was added in
response to a bug report.
However, this is rather expensive, especially when running the tests
under valgrind, CLOBBER_CACHE_ALWAYS etc. It doesn't seem worth it to
keep running the test, so remove it from all branches (14+).
Backpatch-through: 14
Discussion: https://postgr.es/m/2131017.1623451468@sss.pgh.pa.us
Creation of a file with a very long name can create problems on Windows
due to its file path limits. Work around that by creating the file via a
symlink with a shorter name.
Error displayed by buildfarm animal fairywren.o
Backpatch to all live branches
When set, this environment variable was only effective for data
directories but not for all the other temporary files created by
PostgreSQL::Test::Utils. Keeping the temporary files after a successful
run can be useful for debugging purposes.
The documentation is updated to reflect the new behavior, with contents
available in doc/ since v16 and in src/test/perl/README since v15.
Author: Jacob Champion
Reviewed-by: Daniel Gustafsson
Discussion: https://postgr.es/m/CAAWbhmgHtDH1SGZ+Fw05CsXtE0mzTmjbuUxLB9mY9iPKgM6cUw@mail.gmail.com
Discussion: https://postgr.es/m/YyPd9unV14SX2bLF@paquier.xyz
Backpatch-through: 11
Commit f24523672d fixed a memory leak by moving the modifiedCols bitmap
into the per-row memory context. In the case of AFTER UPDATE triggers,
the bitmap is however referenced from an event kept until the end of the
query, resulting in a use-after-free bug.
Fixed by copying the bitmap into the AfterTriggerEvents memory context,
which is the one where we keep the trigger events. There's only one
place that needs to do the copy, but the memory context may not exist
yet. Doing that in a separate function seems more readable.
Report by Alexander Pyhalov, fix by me. Backpatch to 13, where the
bitmap was added to the event by commit 71d60e2aa0.
Reported-by: Alexander Pyhalov
Backpatch-through: 13
Discussion: https://postgr.es/m/acddb17c89b0d6cb940eaeda18c08bbe@postgrespro.ru
The Incremental Sort had a couple issues, resulting in leaking memory
during rescans, possibly triggering OOM. The code had a couple of
related flaws:
1. During rescans, the sort states were reset but then also set to NULL
(despite the comment saying otherwise). ExecIncrementalSort then
sees NULL and initializes a new sort state, leaking the memory used
by the old one.
2. Initializing the sort state also automatically rebuilt the info about
presorted keys, leaking the already initialized info. presorted_keys
was also unnecessarily reset to NULL.
Patch by James Coleman, based on patches by Laurenz Albe and Tom Lane.
Backpatch to 13, where Incremental Sort was introduced.
Author: James Coleman, Laurenz Albe, Tom Lane
Reported-by: Laurenz Albe, Zu-Ming Jiang
Backpatch-through: 13
Discussion: https://postgr.es/m/b2bd02dff61af15e3526293e2771f874cf2a3be7.camel%40cybertec.at
Discussion: https://postgr.es/m/db03c582-086d-e7cd-d4a1-3bc722f81765%40inf.ethz.ch
The logic that introduced partitioned indexes missed a few things when
invalidating a partitioned index when these are created, still the code
is written to handle recursions:
1) If created from scratch because a mapping index could not be found,
the new index created could be itself invalid, if for example it was a
partitioned index with one of its leaves invalid.
2) A CCI was missing when indisvalid is set for a parent index, leading
to inconsistent trees when recursing across more than one level for a
partitioned index creation if an invalidation of the parent was
required.
This could lead to the creation of a partition index tree where some of
the partitioned indexes are marked as invalid, but some of the parents
are marked valid, which is not something that should happen (as
validatePartitionedIndex() defines, indisvalid is switched to true for a
partitioned index iff all its partitions are themselves valid).
This patch makes sure that indisvalid is set to false on a partitioned
index if at least one of its partition is invalid. The flag is set to
true if *all* its partitions are valid.
The regression test added in this commit abuses of a failed concurrent
index creation, marked as invalid, that maps with an index created on
its partitioned table afterwards.
Reported-by: Alexander Lakhin
Reviewed-by: Alexander Lakhin
Discussion: https://postgr.es/m/14987634-43c0-0cb3-e075-94d423607e08@gmail.com
Backpatch-through: 11
If the given composite datum is toasted out-of-line,
DatumGetHeapTupleHeader will perform database accesses to detoast it.
That can invalidate the result of get_cached_rowtype, as documented
(perhaps not plainly enough) in that function's API spec; which leads
to strange errors or crashes when we try to use the TupleDesc to read
the tuple. In short then, trying to update a field of a composite
column could fail intermittently if the overall column value is wide
enough to require toasting.
We can fix the bug at no cost by just changing the order of
operations, since we don't need the TupleDesc until after detoasting.
(Other callers of get_cached_rowtype appear to get this right already,
so there's only one bug.)
Note that the added regression test case reveals this bug reliably
only with debug_discard_caches/CLOBBER_CACHE_ALWAYS.
Per bug #17994 from Alexander Lakhin. Sadly, this patch does not fix
the missing-values issue revealed in the bug discussion; we'll need
some more work to cover that.
Discussion: https://postgr.es/m/17994-5c7100b51b4790e9@postgresql.org
It was walking into the ColumnDef->compression field, which is not a
node but a string. This code is currently not reachable (because the
compression field is only set in situations that don't go through
raw_expression_tree_walker()), but if it had been, this could have
behaved erratically.
A portion of ALTER TABLE .. ATTACH PARTITION is to ensure that the
partition being attached to the partitioned table has a correct set of
indexes, so as there is a consistent index mapping between the
partitioned table and its new-to-be partition. However, as introduced
in 8b08f7d, the current logic could choose an invalid index as a match,
which is something that can exist when dealing with more than two levels
of partitioning, like attaching a partitioned table (that has
partitions, with an index created by CREATE INDEX ON ONLY) to another
partitioned table.
A partitioned index with indisvalid set to false is equivalent to an
incomplete partition tree, meaning that an invalid partitioned index
does not have indexes defined in all its partitions. Hence, choosing an
invalid partitioned index can create inconsistent partition index trees,
where the parent attaching to is valid, but its partition may be
invalid.
In the report from Alexander Lakhin, this showed up as an assertion
failure when validating an index. Without assertions enabled, the
partition index tree would be actually broken, as indisvalid should
be switched to true for a partitioned index once all its partitions are
themselves valid. With two levels of partitioning, the top partitioned
table used a valid index and was able to link to an invalid index stored
on its partition, itself a partitioned table.
I have studied a few options here (like the possibility to switch
indisvalid to false for the parent), but came down to the conclusion
that we'd better rely on a simple rule: invalid indexes had better never
be chosen, so as the partition attached uses and creates indexes that
the parent expects. Some regression tests are added to provide some
coverage. Note that the existing coverage is not impacted.
This is a problem since partitioned indexes exist, so backpatch all the
way down to v11.
Reported-by: Alexander Lakhin
Discussion: https://postgr.es/14987634-43c0-0cb3-e075-94d423607e08@gmail.com
Backpatch-through: 11
TParserGet() recurses for some token types, meaning it's possible
to drive it to stack overflow. Since this is a minority behavior,
I chose to add the check_stack_depth() call to the two places that
recurse rather than doing it during every single call.
While at it, add CHECK_FOR_INTERRUPTS(), because this can run
unpleasantly long for long inputs.
Per bug #17995 from Zuming Jiang. This is old, so back-patch
to all supported branches.
Discussion: https://postgr.es/m/17995-9f20ff3e6389db4c@postgresql.org
One of the tests for the pipeline mode with portal description expects a
non-NULL PQgetResult, but used an incorrect error message on failure,
telling that PQgetResult being NULL was the expected result.
Author: Jelte Fennema
Discussion: https://postgr.es/m/CAGECzQTkShHecFF+EZrm94Lbsu2ej569T=bz+PjMbw9Aiioxuw@mail.gmail.com
Backpatch-through: 14
By default, triggers and rules do not fire on a logical replication
subscriber based on the "session_replication_role" GUC being set to
"replica". However, the docs in the logical replication section assumed
that the reader understood how this GUC worked. This modifies the docs to
be more explicit and links back to the GUC itself.
Author: Jonathan Katz, Peter Smith
Reviewed-by: Vignesh C, Euler Taveira
Backpatch-through: 11
Discussion: https://postgr.es/m/5bb2c9a2-499f-e1a2-6e33-5ce96b35cc4a@postgresql.org
Statistics defined by the CREATE STATISTICS command are only used to
assist with the selectivity estimations of base relations, never for
joins. Here we mention this fact in the notes section of the CREATE
STATISTICS command.
Discussion: https://postgr.es/m/CAApHDvrMuVgDOrmg_EtFDZ=AOovq6EsJNnHH1ddyZ8EqL4yzMw@mail.gmail.com
Backpatch-through: 11
Avoid "right sibling %u of block %u is not next child" errors when
vacuuming a corrupt nbtree index. Just LOG the issue and press on.
That way VACUUM will have a decent chance of finishing off all required
processing for the index (and for the table as a whole).
This is similar to recent work from commit 5abff197, as well as work
from commit 5b861baa (later backpatched as commit 43e409ce), which
taught nbtree VACUUM to keep going when its "re-find" check fails. The
hardening added by this commit takes place directly after the "re-find"
check, right before the critical section for the first stage of page
deletion.
Author: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/CAH2-Wz=dayg0vjs4+er84TS9ami=csdzjpuiCGbEw=idhwqhzQ@mail.gmail.com
Backpatch: 11- (all supported versions).
exec_parse_message() wants to create a cached plan in all cases,
including for empty input. The empty-input path does not have
a test for being in an aborted transaction, making it possible
that plancache.c will fail due to trying to do database lookups
even though there's no real work to do.
One solution would be to throw an aborted-transaction error in
this path too, but it's not entirely clear whether the lack of
such an error was intentional or whether some clients might be
relying on non-error behavior. Instead, let's hack plancache.c
so that it treats empty statements with the same logic it
already had for transaction control commands, ensuring that it
can soldier through even in an already-aborted transaction.
Per bug #17983 from Alexander Lakhin. Back-patch to all
supported branches.
Discussion: https://postgr.es/m/17983-da4569fcb878672e@postgresql.org
This partially reverts 68cb5af, as using archiving to enforce the
rename of the last partial segment of the old timeline at promotion to
use .partial as suffix is impacting the tests when it does switchovers.
As showed by the logs gathered by the CI in the tests that failed, a new
standby may fail to find the WAL segment it needs to follow a promoted
instance with its timeline jump, as it got renamed to .partial.
This problem would manifest as a run timeout with 009_twophase.pl, as
the new standby repeatedly requests a segment from the promoted primary
that it would not find.
Reported-by: Nathan Bossart
Discussion: https://postgr.es/m/20230621043345.GA787473@nathanxps13
Backpatch-through: 13
The existing errhint message and docs were missing the fact that we can't
disassociate from the slot unless the subscription is disabled.
Author: Robert Sjöblom, Peter Smith
Reviewed-by: Peter Eisentraut, Amit Kapila
Backpatch-through: 11
Discussion: https://postgr.es/m/807bdf85-61ea-88e2-5712-6d9fcd4eabff@fortnox.se
If the inner-side expressions contain PARAM_EXEC Params, we must
re-hash whenever the values of those Params change. The executor
mechanism for that exists already, but we failed to invoke it because
finalize_plan() neglected to search the Hash.hashkeys field for
Params. This allowed a previous scan's hash table to be re-used
when it should not be, leading to rows missing from the join's output.
(I believe incorrectly-included join rows are impossible however,
since checking the real hashclauses would reject false matches.)
This bug is very ancient, dating probably to d24d75ff1 of 7.4.
Sadly, this simple fix depends on the plan representational changes
made by 2abd7ae9b, so it will only work back to v12. I thought
about trying to make some kind of hack for v11, but I'm leery
of putting code significantly different from what is used in the
newer branches into a nearly-EOL branch. Seeing that the bug
escaped detection for a full twenty years, problematic cases
must be rare; so I don't feel too awful about leaving v11 as-is.
Per bug #17985 from Zuming Jiang. Back-patch to v12.
Discussion: https://postgr.es/m/17985-748b66607acd432e@postgresql.org
This is a follow-up of f663b00, that has been committed to v13 and v14,
tweaking the TAP test for two-phase transactions so as it provides
coverage for the bug that has been fixed. This change is done in its
own commit for clarity, as v15 and HEAD did not show the problematic
behavior, still missed coverage for it.
While on it, this adds a comment about the dependency of the last
partial segment rename and RecoverPreparedTransactions() at the end of
recovery, as that can be easy to miss.
Author: Michael Paquier
Reviewed-by: Kyotaro Horiguchi
Discussion: https://postgr.es/m/743b9b45a2d4013bd90b6a5cba8d6faeb717ee34.camel@cybertec.at
Backpatch-through: 13
When archiving is enabled, a promotion request would fail with the
following error when some 2PC transaction needs to be recovered from
WAL, preventing the promotion to complete:
FATAL: requested WAL segment pg_wal/000000010000000000000001 has already been removed
The origin of the problem is that the last partial segment of the old
timeline is renamed before recovering the 2PC data via
RecoverPreparedTransactions() at the end of recovery, causing the FATAL
because the segment wanted is now renamed with a .partial suffix. This
commit reorders a bit the end-of-recovery actions so as the execution of
recovery_end_command, the cleanup of the old segments of the old
timeline (RemoveNonParentXlogFiles) and the last partial segment rename
are done after the 2PC transaction data is recovered with
RecoverPreparedTransactions(). This makes the order of these
end-of-recovery actions more consistent with ~15, at the exception of
the end-of-recovery checkpoint that still needs to happen before all the
actions reordered here in v13 and v14, contrary to what 15~ does.
v15 and newer versions have "fixed" this problem somewhat accidentally
with 811051c, where the end-of-recovery actions got reordered. In this
case, the recovery of 2PC transactions happens before the renaming of
the last partial segment of the old timeline.
v13 and v14 are the versions that can easily see this problem as per the
refactoring of 38a95731 where XLogReaderState is reset in
XLogBeginRead() before reading the 2PC transaction data. v11 and v12
could also see this problem, but may finish by reading the 2PC data from
some of the WAL buffers instead. Perhaps something could be done for
these two branches, but I am not really excited about doing something on
these per the lack of complaints and per the fact that v11 is soon going
to be EOL'd soon (there is always a risk of breaking something).
Note that the TAP test 009_twophase.pl is able to exhibit the issue if
it enables archiving on the primary node, which does not impact the test
coverage as restore_command would remain unused. This is something that
should be changed on v15 and HEAD as well, so this will be changed in a
separate commit for clarity.
Author: Julian Markwort
Reviewed-by: Kyotaro Horiguchi, Michael Paquier
Discussion: https://postgr.es/m/743b9b45a2d4013bd90b6a5cba8d6faeb717ee34.camel@cybertec.at
Backpatch-through: 13
Here we adjust relation_has_unique_index_for() so that it no longer makes
use of partial unique indexes as uniqueness proofs. It is incorrect to
use these as the predicates used by check_index_predicates() to set
predOK makes use of not only baserestrictinfo quals as proofs, but also
qual from join conditions. For relation_has_unique_index_for()'s case, we
need to know the relation is unique for a given set of columns before any
joins are evaluated, so if predOK was only set to true due to some join
qual, then it's unsafe to use such indexes in
relation_has_unique_index_for(). The final plan may not even make use
of that index, which could result in reading tuples that are not as
unique as the planner previously expected them to be.
Bug: #17975
Reported-by: Tor Erik Linnerud
Backpatch-through: 11, all supported versions
Discussion: https://postgr.es/m/17975-98a90c156f25c952%40postgresql.org
Michael Paquier
Andres Freund
Tom Lane
David Rowley
Andrew Dunstan
Heikki Linnakangas
Amit Kapila
Peter Eisentraut
Thomas Munro
Tomas Vondra
Bruce Momjian
Peter Geoghegan