Crash recovery on standby may encounter missing directories when
replaying create database WAL records. Prior to this patch, the standby
would fail to recover in such a case. However, the directories could be
legitimately missing. Consider a sequence of WAL records as follows:
CREATE DATABASE
DROP DATABASE
DROP TABLESPACE
If, after replaying the last WAL record and removing the tablespace
directory, the standby crashes and has to replay the create database
record again, the crash recovery must be able to move on.
This patch adds a mechanism similar to invalid-page tracking, to keep a
tally of missing directories during crash recovery. If all the missing
directory references are matched with corresponding drop records at the
end of crash recovery, the standby can safely continue following the
primary.
Backpatch to 13, at least for now. The bug is older, but fixing it in
older branches requires more careful study of the interactions with
commit e6d8069522, which appeared in 13.
A new TAP test file is added to verify the condition. However, because
it depends on commit d6d317dbf6, it can only be added to branch
master. I (Álvaro) manually verified that the code behaves as expected
in branch 14. It's a bit nervous-making to leave the code uncovered by
tests in older branches, but leaving the bug unfixed is even worse.
Also, the main reason this fix took so long is precisely that we
couldn't agree on a good strategy to approach testing for the bug, so
perhaps this is the best we can do.
Diagnosed-by: Paul Guo <paulguo@gmail.com>
Author: Paul Guo <paulguo@gmail.com>
Author: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
Author: Asim R Praveen <apraveen@pivotal.io>
Discussion: https://postgr.es/m/CAEET0ZGx9AvioViLf7nbR_8tH9-=27DN5xWJ2P9-ROH16e4JUA@mail.gmail.com
If TRUNCATE causes some buffers to be invalidated and thus the
checkpoint does not flush them, TRUNCATE must also ensure that the
corresponding files are truncated on disk. Otherwise, a replay
from the checkpoint might find that the buffers exist but have
the wrong contents, which may cause replay to fail.
Report by Teja Mupparti. Patch by Kyotaro Horiguchi, per a design
suggestion from Heikki Linnakangas, with some changes to the
comments by me. Review of this and a prior patch that approached
the issue differently by Heikki Linnakangas, Andres Freund, Álvaro
Herrera, Masahiko Sawada, and Tom Lane.
Discussion: http://postgr.es/m/BYAPR06MB6373BF50B469CA393C614257ABF00@BYAPR06MB6373.namprd06.prod.outlook.com
If we run out of space in the checkpointer sync request queue (which is
hopefully rare on real systems, but common with very small buffer pool),
we wait for it to drain. While waiting, we should report that as a wait
event so that users know what is going on, and also handle postmaster
death, since otherwise the loop might never terminate if the
checkpointer has exited.
Back-patch to 12. Although the problem exists in earlier releases too,
the code is structured differently before 12 so I haven't gone any
further for now, in the absence of field complaints.
Reported-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20220226213942.nb7uvb2pamyu26dj%40alap3.anarazel.de
The checkpointer shouldn't ignore its latch. Other backends may be
waiting for it to drain the request queue. Hopefully real systems don't
have a full queue often, but the condition is reached easily when
shared_buffers is small.
This involves defining a new wait event, which will appear in the
pg_stat_activity view often due to spread checkpoints.
Back-patch only to 14. Even though the problem exists in earlier
branches too, it's hard to hit there. In 14 we stopped using signal
handlers for latches on Linux, *BSD and macOS, which were previously
hiding this problem by interrupting the sleep (though not reliably, as
the signal could arrive before the sleep begins; precisely the problem
latches address).
Reported-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20220226213942.nb7uvb2pamyu26dj%40alap3.anarazel.de
This macro cast the result to BlockNumber after shifting, not before,
which is the wrong thing. Per the C spec, the uint16 fields would
promote to int not unsigned int, so that (for 32-bit int) the shift
potentially shifts a nonzero bit into the sign position. I doubt
there are any production systems where this would actually end with
the wrong answer, but it is undefined behavior per the C spec, and
clang's -fsanitize=undefined option reputedly warns about it on some
platforms. (I can't reproduce that right now, but the code is
undeniably wrong per spec.) It's easy to fix by casting to
BlockNumber (uint32) in the proper places.
It's been wrong for ages, so back-patch to all supported branches.
Report and patch by Zhihong Yu (cosmetic tweaking by me)
Discussion: https://postgr.es/m/CALNJ-vT9r0DSsAOw9OXVJFxLENoVS_68kJ5x0p44atoYH+H4dg@mail.gmail.com
"regress" is a new mode added to compute_query_id aimed at facilitating
regression testing when a module computing query IDs is loaded into the
backend, like pg_stat_statements. It works the same way as "auto",
meaning that query IDs are computed if a module enables it, except that
query IDs are hidden in EXPLAIN outputs to ensure regression output
stability.
Like any GUCs of the kind (force_parallel_mode, etc.), this new
configuration can be added to an instance's postgresql.conf, or just
passed down with PGOPTIONS at command level. compute_query_id uses an
enum for its set of option values, meaning that this addition ensures
ABI compatibility.
Using this new configuration mode allows installcheck-world to pass when
running the tests on an instance with pg_stat_statements enabled,
stabilizing the test output while checking the paths doing query ID
computations.
Reported-by: Anton Melnikov
Reviewed-by: Julien Rouhaud
Discussion: https://postgr.es/m/1634283396.372373993@f75.i.mail.ru
Discussion: https://postgr.es/m/YgHlxgc/OimuPYhH@paquier.xyz
Backpatch-through: 14
GCC 12 complains that set_stack_base is storing the address of
a local variable in a long-lived pointer. This is an entirely
reasonable warning (indeed, it just helped us find a bug);
but that behavior is intentional here. We can work around it
by using __builtin_frame_address(0) instead of a specific local
variable; that produces an address a dozen or so bytes different,
in my testing, but we don't care about such a small difference.
Maybe someday a compiler lacking that function will start to issue
a similar warning, but we'll worry about that when it happens.
Patch by me, per a suggestion from Andres Freund. Back-patch to
v12, which is as far back as the patch will go without some pain.
(Recently-established project policy would permit a back-patch as
far as 9.2, but I'm disinclined to expend the work until GCC 12
is much more widespread.)
Discussion: https://postgr.es/m/3773792.1645141467@sss.pgh.pa.us
Commit 6a2a70a02 supposed that any platform having <sys/epoll.h>
would also have <sys/signalfd.h>. It turns out there are still a
few people using platforms where that's not so, so we'd better make
a separate configure probe for it. But since it took this long to
notice, I'm content with the decision to not have a separate code
path for epoll-only machines; we'll just fall back to using poll()
for these stragglers.
Per gripe from Gabriela Serventi. Back-patch to v14 where this
code came in.
Discussion: https://postgr.es/m/CAHOHWE-JjJDfcYuLAAEO7Jk07atFAU47z8TzHzg71gbC0aMy=g@mail.gmail.com
Although select_common_type() has a failure-return convention, an
apparent successful return just provides a type OID that *might* work
as a common supertype; we've not validated that the required casts
actually exist. In the mainstream use-cases that doesn't matter,
because we'll proceed to invoke coerce_to_common_type() on each input,
which will fail appropriately if the proposed common type doesn't
actually work. However, a few callers didn't read the (nonexistent)
fine print, and thought that if they got back a nonzero OID then the
coercions were sure to work.
This affects in particular the recently-added "anycompatible"
polymorphic types; we might think that a function/operator using
such types matches cases it really doesn't. A likely end result
of that is unexpected "ambiguous operator" errors, as for example
in bug #17387 from James Inform. Another, much older, case is that
the parser might try to transform an "x IN (list)" construct to
a ScalarArrayOpExpr even when the list elements don't actually have
a common supertype.
It doesn't seem desirable to add more checking to select_common_type
itself, as that'd just slow down the mainstream use-cases. Instead,
write a separate function verify_common_type that performs the
missing checks, and add a call to that where necessary. Likewise add
verify_common_type_from_oids to go with select_common_type_from_oids.
Back-patch to v13 where the "anycompatible" types came in. (The
symptom complained of in bug #17387 doesn't appear till v14, but
that's just because we didn't get around to converting || to use
anycompatible till then.) In principle the "x IN (list)" fix could
go back all the way, but I'm not currently convinced that it makes
much difference in real-world cases, so I won't bother for now.
Discussion: https://postgr.es/m/17387-5dfe54b988444963@postgresql.org
Commit 8431e296ea reworked ProcArrayApplyRecoveryInfo to sort XIDs
before adding them to KnownAssignedXids. But the XIDs are sorted using
xidComparator, which compares the XIDs simply as uint32 values, not
logically. KnownAssignedXidsAdd() however expects XIDs in logical order,
and calls TransactionIdFollowsOrEquals() to enforce that. If there are
XIDs for which the two orderings disagree, an error is raised and the
recovery fails/restarts.
Hitting this issue is fairly easy - you just need two transactions, one
started before the 4B limit (e.g. XID 4294967290), the other sometime
after it (e.g. XID 1000). Logically (4294967290 <= 1000) but when
compared using xidComparator we try to add them in the opposite order.
Which makes KnownAssignedXidsAdd() fail with an error like this:
ERROR: out-of-order XID insertion in KnownAssignedXids
This only happens during replica startup, while processing RUNNING_XACTS
records to build the snapshot. Once we reach STANDBY_SNAPSHOT_READY, we
skip these records. So this does not affect already running replicas,
but if you restart (or create) a replica while there are transactions
with XIDs for which the two orderings disagree, you may hit this.
Long-running transactions and frequent replica restarts increase the
likelihood of hitting this issue. Once the replica gets into this state,
it can't be started (even if the old transactions are terminated).
Fixed by sorting the XIDs logically - this is fine because we're dealing
with normal XIDs (because it's XIDs assigned to backends) and from the
same wraparound epoch (otherwise the backends could not be running at
the same time on the primary node). So there are no problems with the
triangle inequality, which is why xidComparator compares raw values.
Investigation and root cause analysis by Abhijit Menon-Sen. Patch by me.
This issue is present in all releases since 9.4, however releases up to
9.6 are EOL already so backpatch to 10 only.
Reviewed-by: Abhijit Menon-Sen
Reviewed-by: Alvaro Herrera
Backpatch-through: 10
Discussion: https://postgr.es/m/36b8a501-5d73-277c-4972-f58a4dce088a%40enterprisedb.com
In logical replication mode, a WalSender is supposed to be able
to execute any regular SQL command, as well as the special
replication commands. Poor design of the replication-command
parser caused it to fail in various cases, notably:
* semicolons embedded in a command, or multiple SQL commands
sent in a single message;
* dollar-quoted literals containing odd numbers of single
or double quote marks;
* commands starting with a comment.
The basic problem here is that we're trying to run repl_scanner.l
across the entire input string even when it's not a replication
command. Since repl_scanner.l does not understand all of the
token types known to the core lexer, this is doomed to have
failure modes.
We certainly don't want to make repl_scanner.l as big as scan.l,
so instead rejigger stuff so that we only lex the first token of
a non-replication command. That will usually look like an IDENT
to repl_scanner.l, though a comment would end up getting reported
as a '-' or '/' single-character token. If the token is a replication
command keyword, we push it back and proceed normally with repl_gram.y
parsing. Otherwise, we can drop out of exec_replication_command()
without examining the rest of the string.
(It's still theoretically possible for repl_scanner.l to fail on
the first token; but that could only happen if it's an unterminated
single- or double-quoted string, in which case you'd have gotten
largely the same error from the core lexer too.)
In this way, repl_gram.y isn't involved at all in handling general
SQL commands, so we can get rid of the SQLCmd node type. (In
the back branches, we can't remove it because renumbering enum
NodeTag would be an ABI break; so just leave it sit there unused.)
I failed to resist the temptation to clean up some other sloppy
coding in repl_scanner.l while at it. The only externally-visible
behavior change from that is it now accepts \r and \f as whitespace,
same as the core lexer.
Per bug #17379 from Greg Rychlewski. Back-patch to all supported
branches.
Discussion: https://postgr.es/m/17379-6a5c6cfb3f1f5e77@postgresql.org
This reverts commits ab27df2, af8d530 and 3a0cced, that introduced
pg_cryptohash_error(). In order to make the core code able to pass down
the new error types that this introduced, some of the MD5-related
routines had to be reworked, causing an ABI breakage, but we found that
some external extensions rely on them. Maintaining compatibility
outweights the error report benefits, so just revert the change in v14.
Reported-by: Laurenz Albe
Discussion: https://postgr.es/m/9f0c0a96d28cf14fc87296bbe67061c14eb53ae8.camel@cybertec.at
The existing cryptohash facility was causing problems in some code paths
related to MD5 (frontend and backend) that relied on the fact that the
only type of error that could happen would be an OOM, as the MD5
implementation used in PostgreSQL ~13 (the in-core implementation is
used when compiling with or without OpenSSL in those older versions),
could fail only under this circumstance.
The new cryptohash facilities can fail for reasons other than OOMs, like
attempting MD5 when FIPS is enabled (upstream OpenSSL allows that up to
1.0.2, Fedora and Photon patch OpenSSL 1.1.1 to allow that), so this
would cause incorrect reports to show up.
This commit extends the cryptohash APIs so as callers of those routines
can fetch more context when an error happens, by using a new routine
called pg_cryptohash_error(). The error states are stored within each
implementation's internal context data, so as it is possible to extend
the logic depending on what's suited for an implementation. The default
implementation requires few error states, but OpenSSL could report
various issues depending on its internal state so more is needed in
cryptohash_openssl.c, and the code is shaped so as we are always able to
grab the necessary information.
The core code is changed to adapt to the new error routine, painting
more "const" across the call stack where the static errors are stored,
particularly in authentication code paths on variables that provide
log details. This way, any future changes would warn if attempting to
free these strings. The MD5 authentication code was also a bit blurry
about the handling of "logdetail" (LOG sent to the postmaster), so
improve the comments related that, while on it.
The origin of the problem is 87ae969, that introduced the centralized
cryptohash facility. Extra changes are done for pgcrypto in v14 for the
non-OpenSSL code path to cope with the improvements done by this
commit.
Reported-by: Michael Mühlbeyer
Author: Michael Paquier
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/89B7F072-5BBE-4C92-903E-D83E865D9367@trivadis.com
Backpatch-through: 14
Commit 4ace45677 failed to fix the problem fully, because the
same issue of attempting to fetch a non-returnable index column
can occur when rechecking the indexqual after using a lossy index
operator. Moreover, it broke EXPLAIN for such indexquals (which
indicates a gap in our test cases :-().
Revert the code changes of 4ace45677 in favor of adding a new field
to struct IndexOnlyScan, containing a version of the indexqual that
can be executed against the index-returned tuple without using any
non-returnable columns. (The restrictions imposed by check_index_only
guarantee this is possible, although we may have to recompute indexed
expressions.) Support construction of that during setrefs.c
processing by marking IndexOnlyScan.indextlist entries as resjunk
if they can't be returned, rather than removing them entirely.
(We could alternatively require setrefs.c to look up the IndexOptInfo
again, but abusing resjunk this way seems like a reasonably safe way
to avoid needing to do that.)
This solution isn't great from an API-stability standpoint: if there
are any extensions out there that build IndexOnlyScan structs directly,
they'll be broken in the next minor releases. However, only a very
invasive extension would be likely to do such a thing. There's no
change in the Path representation, so typical planner extensions
shouldn't have a problem.
As before, back-patch to all supported branches.
Discussion: https://postgr.es/m/3179992.1641150853@sss.pgh.pa.us
Discussion: https://postgr.es/m/17350-b5bdcf476e5badbb@postgresql.org
If an index has both returnable and non-returnable columns, and one of
the non-returnable columns is an expression using a Var that is in a
returnable column, then a query returning that expression could result
in an index-only scan plan that attempts to read the non-returnable
column, instead of recomputing the expression from the returnable
column as intended.
To fix, redefine the "indextlist" list of an IndexOnlyScan plan node
as containing null Consts in place of any non-returnable columns.
This solves the problem by preventing setrefs.c from falsely matching
to such entries. The executor is happy since it only cares about the
exposed types of the entries, and ruleutils.c doesn't care because a
correct plan won't reference those entries. I considered some other
ways to prevent setrefs.c from doing the wrong thing, but this way
seems good since (a) it allows a very localized fix, (b) it makes
the indextlist structure more compact in many cases, and (c) the
indextlist is now a more faithful representation of what the index AM
will actually produce, viz. nulls for any non-returnable columns.
This is easier to hit since we introduced included columns, but it's
possible to construct failing examples without that, as per the
added regression test. Hence, back-patch to all supported branches.
Per bug #17350 from Louis Jachiet.
Discussion: https://postgr.es/m/17350-b5bdcf476e5badbb@postgresql.org
catalog/pg_class.h was stating that REPLICA_IDENTITY_INDEX with a
dropped index is equivalent to REPLICA_IDENTITY_DEFAULT. The code tells
a different story, as it is equivalent to REPLICA_IDENTITY_NOTHING.
The behavior exists since the introduction of replica identities, and
fe7fd4e even added tests for this case but I somewhat forgot to fix this
comment.
While on it, this commit reorganizes the documentation about replica
identities on the ALTER TABLE page, and a note is added about the case
of dropped indexes with REPLICA_IDENTITY_INDEX.
Author: Michael Paquier, Wei Wang
Reviewed-by: Euler Taveira
Discussion: https://postgr.es/m/OS3PR01MB6275464AD0A681A0793F56879E759@OS3PR01MB6275.jpnprd01.prod.outlook.com
Backpatch-through: 10
Surround the contents with a test that the feature is enabled by
configure, to silence header checking tools on systems without GSSAPI
installed.
Backpatch to 12, where the file appeared.
Discussion: https://postgr.es/m/202111161709.u3pbx5lxdimt@alvherre.pgsql
It's possible that a subplan below a Memoize node contains a parameter
from above the Memoize node. If this parameter changes then cache entries
may become out-dated due to the new parameter value.
Previously Memoize was mistakenly not aware of this. We fix this here by
flushing the cache whenever a parameter that's not part of the cache
key changes.
Bug: #17213
Reported by: Elvis Pranskevichus
Author: David Rowley
Discussion: https://postgr.es/m/17213-988ed34b225a2862@postgresql.org
Backpatch-through: 14, where Memoize was added
It's possible that a subplan below a Memoize node contains a parameter
from above the Memoize node. If this parameter changes then cache entries
may become out-dated due to the new parameter value.
Previously Memoize was mistakenly not aware of this. We fix this here by
flushing the cache whenever a parameter that's not part of the cache
key changes.
Bug: #17213
Reported by: Elvis Pranskevichus
Author: David Rowley
Discussion: https://postgr.es/m/17213-988ed34b225a2862@postgresql.org
Backpatch-through: 14, where Memoize was added
Memoize would always use the hash equality operator for the cache key
types to determine if the current set of parameters were the same as some
previously cached set. Certain types such as floating points where -0.0
and +0.0 differ in their binary representation but are classed as equal by
the hash equality operator may cause problems as unless the join uses the
same operator it's possible that whichever join operator is being used
would be able to distinguish the two values. In which case we may
accidentally return in the incorrect rows out of the cache.
To fix this here we add a binary mode to Memoize to allow it to the
current set of parameters to previously cached values by comparing
bit-by-bit rather than logically using the hash equality operator. This
binary mode is always used for LATERAL joins and it's used for normal
joins when any of the join operators are not hashable.
Reported-by: Tom Lane
Author: David Rowley
Discussion: https://postgr.es/m/3004308.1632952496@sss.pgh.pa.us
Backpatch-through: 14, where Memoize was added
While determining xid horizons, we skip over backends that are running
Vacuum. We also ignore Create Index Concurrently, or Reindex Concurrently
for the purposes of computing Xmin for Vacuum. But we were not setting the
flags corresponding to these operations when they are performed in
parallel which was preventing Xid horizon from advancing.
The optimization related to skipping Create Index Concurrently, or Reindex
Concurrently operations was implemented in PG-14 but the fix is the same
for the Parallel Vacuum as well so back-patched till PG-13.
Author: Masahiko Sawada
Reviewed-by: Amit Kapila
Backpatch-through: 13
Discussion: https://postgr.es/m/CAD21AoCLQqgM1sXh9BrDFq0uzd3RBFKi=Vfo6cjjKODm0Onr5w@mail.gmail.com
The server collects up to a bufferload of data whenever it reads data
from the client socket. When SSL or GSS encryption is requested
during startup, any additional data received with the initial
request message remained in the buffer, and would be treated as
already-decrypted data once the encryption handshake completed.
Thus, a man-in-the-middle with the ability to inject data into the
TCP connection could stuff some cleartext data into the start of
a supposedly encryption-protected database session.
This could be abused to send faked SQL commands to the server,
although that would only work if the server did not demand any
authentication data. (However, a server relying on SSL certificate
authentication might well not do so.)
To fix, throw a protocol-violation error if the internal buffer
is not empty after the encryption handshake.
Our thanks to Jacob Champion for reporting this problem.
Security: CVE-2021-23214
Commit b4af70cb, which simplified state managed by VACUUM, performed
refactoring of parallel VACUUM in passing. Confusion about the exact
details of the tasks that the leader process is responsible for led to
code that made it possible for parallel VACUUM to miss a subset of the
table's indexes entirely. Specifically, indexes that fell under the
min_parallel_index_scan_size size cutoff were missed. These indexes are
supposed to be vacuumed by the leader (alongside any parallel unsafe
indexes), but weren't vacuumed at all. Affected indexes could easily
end up with duplicate heap TIDs, once heap TIDs were recycled for new
heap tuples. This had generic symptoms that might be seen with almost
any index corruption involving structural inconsistencies between an
index and its table.
To fix, make sure that the parallel VACUUM leader process performs any
required index vacuuming for indexes that happen to be below the size
cutoff. Also document the design of parallel VACUUM with these
below-size-cutoff indexes.
It's unclear how many users might be affected by this bug. There had to
be at least three indexes on the table to hit the bug: a smaller index,
plus at least two additional indexes that themselves exceed the size
cutoff. Cases with just one additional index would not run into
trouble, since the parallel VACUUM cost model requires two
larger-than-cutoff indexes on the table to apply any parallel
processing. Note also that autovacuum was not affected, since it never
uses parallel processing.
Test case based on tests from a larger patch to test parallel VACUUM by
Masahiko Sawada.
Many thanks to Kamigishi Rei for her invaluable help with tracking this
problem down.
Author: Peter Geoghegan <pg@bowt.ie>
Author: Masahiko Sawada <sawada.mshk@gmail.com>
Reported-By: Kamigishi Rei <iijima.yun@koumakan.jp>
Reported-By: Andrew Gierth <andrew@tao11.riddles.org.uk>
Diagnosed-By: Andres Freund <andres@anarazel.de>
Bug: #17245
Discussion: https://postgr.es/m/17245-ddf06aaf85735f36@postgresql.org
Discussion: https://postgr.es/m/20211030023740.qbnsl2xaoh2grq3d@alap3.anarazel.de
Backpatch: 14-, where the refactoring commit appears.
As in commits 6301c3ada and e9d9ba2a4, avoid doing repetitive
list_delete_first() operations, since that would be expensive when
there are many files waiting to be unlinked. This is a slightly
larger change than in those cases. We have to keep the list state
valid for calls to AbsorbSyncRequests(), so it's necessary to invent a
"canceled" field instead of immediately deleting PendingUnlinkEntry
entries. Also, because we might not be able to process all the
entries, we need a new list primitive list_delete_first_n().
list_delete_first_n() is almost list_copy_tail(), but it modifies the
input List instead of making a new copy. I found a couple of existing
uses of the latter that could profitably use the new function. (There
might be more, but the other callers look like they probably shouldn't
overwrite the input List.)
As before, back-patch to v13.
Discussion: https://postgr.es/m/CD2F0E7F-9822-45EC-A411-AE56F14DEA9F@amazon.com
The purpose of commit 8a54e12a38 was to
fix this, and it sufficed when the PREPARE TRANSACTION completed before
the CIC looked for lock conflicts. Otherwise, things still broke. As
before, in a cluster having used CIC while having enabled prepared
transactions, queries that use the resulting index can silently fail to
find rows. It may be necessary to reindex to recover from past
occurrences; REINDEX CONCURRENTLY suffices. Fix this for future index
builds by making CIC wait for arbitrarily-recent prepared transactions
and for ordinary transactions that may yet PREPARE TRANSACTION. As part
of that, have PREPARE TRANSACTION transfer locks to its dummy PGPROC
before it calls ProcArrayClearTransaction(). Back-patch to 9.6 (all
supported versions).
Andrey Borodin, reviewed (in earlier versions) by Andres Freund.
Discussion: https://postgr.es/m/01824242-AA92-4FE9-9BA7-AEBAFFEA3D0C@yandex-team.ru
CIC and REINDEX CONCURRENTLY assume backends see their catalog changes
no later than each backend's next transaction start. That failed to
hold when a backend absorbed a relevant invalidation in the middle of
running RelationBuildDesc() on the CIC index. Queries that use the
resulting index can silently fail to find rows. Fix this for future
index builds by making RelationBuildDesc() loop until it finishes
without accepting a relevant invalidation. It may be necessary to
reindex to recover from past occurrences; REINDEX CONCURRENTLY suffices.
Back-patch to 9.6 (all supported versions).
Noah Misch and Andrey Borodin, reviewed (in earlier versions) by Andres
Freund.
Discussion: https://postgr.es/m/20210730022548.GA1940096@gust.leadboat.com
The code does not expect sh_error() to return, but the patch
that made this header usable in frontend didn't get that memo.
While here, plaster unlikely() on the tests that decide whether
to invoke sh_error(), and add our standard copyright notice.
Noted by Andres Freund. Back-patch to v13 where this frontend
support came in.
Discussion: https://postgr.es/m/0D54435C-1199-4361-9D74-2FBDCF8EA164@anarazel.de
During a replication slot creation, an ERROR generated in the same
transaction as the one creating a to-be-exported snapshot would have
left the backend in an inconsistent state, as the associated static
export snapshot state was not being reset on transaction abort, but only
on the follow-up command received by the WAL sender that created this
snapshot on replication slot creation. This would trigger inconsistency
failures if this session tried to export again a snapshot, like during
the creation of a replication slot.
Note that a snapshot export cannot happen in a transaction block, so
there is no need to worry resetting this state for subtransaction
aborts. Also, this inconsistent state would very unlikely show up to
users. For example, one case where this could happen is an
out-of-memory error when building the initial snapshot to-be-exported.
Dilip found this problem while poking at a different patch, that caused
an error in this code path for reasons unrelated to HEAD.
Author: Dilip Kumar
Reviewed-by: Michael Paquier, Zhihong Yu
Discussion: https://postgr.es/m/CAFiTN-s0zA1Kj0ozGHwkYkHwa5U0zUE94RSc_g81WrpcETB5=w@mail.gmail.com
Backpatch-through: 9.6
Prior to v14, we insisted that the query in RETURN QUERY be of a type
that returns tuples. (For instance, INSERT RETURNING was allowed,
but not plain INSERT.) That happened indirectly because we opened a
cursor for the query, so spi.c checked SPI_is_cursor_plan(). As a
consequence, the error message wasn't terribly on-point, but at least
it was there.
Commit 2f48ede08 lost this detail. Instead, plain RETURN QUERY
insisted that the query be a SELECT (by checking for SPI_OK_SELECT)
while RETURN QUERY EXECUTE failed to check the query type at all.
Neither of these changes was intended.
The only convenient place to check this in the EXECUTE case is inside
_SPI_execute_plan, because we haven't done parse analysis until then.
So we need to pass down a flag saying whether to enforce that the
query returns tuples. Fortunately, we can squeeze another boolean
into struct SPIExecuteOptions without an ABI break, since there's
padding space there. (It's unlikely that any extensions would
already be using this new struct, but preserving ABI in v14 seems
like a smart idea anyway.)
Within spi.c, it seemed like _SPI_execute_plan's parameter list
was already ridiculously long, and I didn't want to make it longer.
So I thought of passing SPIExecuteOptions down as-is, allowing that
parameter list to become much shorter. This makes the patch a bit
more invasive than it might otherwise be, but it's all internal to
spi.c, so that seems fine.
Per report from Marc Bachmann. Back-patch to v14 where the
faulty code came in.
Discussion: https://postgr.es/m/1F2F75F0-27DF-406F-848D-8B50C7EEF06A@gmail.com
Commit 84f5c2908 forgot to consider the possibility that
EnsurePortalSnapshotExists could run inside a subtransaction with
lifespan shorter than the Portal's. In that case, the new active
snapshot would be popped at the end of the subtransaction, leaving
a dangling pointer in the Portal, with mayhem ensuing.
To fix, make sure the ActiveSnapshot stack entry is marked with
the same subtransaction nesting level as the associated Portal.
It's certainly safe to do so since we won't be here at all unless
the stack is empty; hence we can't create an out-of-order stack.
Let's also apply this logic in the case where PortalRunUtility
sets portalSnapshot, just to be sure that path can't cause similar
problems. It's slightly less clear that that path can't create
an out-of-order stack, so add an assertion guarding it.
Report and patch by Bertrand Drouvot (with kibitzing by me).
Back-patch to v11, like the previous commit.
Discussion: https://postgr.es/m/ff82b8c5-77f4-3fe7-6028-fcf3303e82dd@amazon.com
Physical replication always ships WAL segment files to replicas once
they are complete. This is a problem if one WAL record is split across
a segment boundary and the primary server crashes before writing down
the segment with the next portion of the WAL record: WAL writing after
crash recovery would happily resume at the point where the broken record
started, overwriting that record ... but any standby or backup may have
already received a copy of that segment, and they are not rewinding.
This causes standbys to stop following the primary after the latter
crashes:
LOG: invalid contrecord length 7262 at A8/D9FFFBC8
because the standby is still trying to read the continuation record
(contrecord) for the original long WAL record, but it is not there and
it will never be. A workaround is to stop the replica, delete the WAL
file, and restart it -- at which point a fresh copy is brought over from
the primary. But that's pretty labor intensive, and I bet many users
would just give up and re-clone the standby instead.
A fix for this problem was already attempted in commit 515e3d84a0, but
it only addressed the case for the scenario of WAL archiving, so
streaming replication would still be a problem (as well as other things
such as taking a filesystem-level backup while the server is down after
having crashed), and it had performance scalability problems too; so it
had to be reverted.
This commit fixes the problem using an approach suggested by Andres
Freund, whereby the initial portion(s) of the split-up WAL record are
kept, and a special type of WAL record is written where the contrecord
was lost, so that WAL replay in the replica knows to skip the broken
parts. With this approach, we can continue to stream/archive segment
files as soon as they are complete, and replay of the broken records
will proceed across the crash point without a hitch.
Because a new type of WAL record is added, users should be careful to
upgrade standbys first, primaries later. Otherwise they risk the standby
being unable to start if the primary happens to write such a record.
A new TAP test that exercises this is added, but the portability of it
is yet to be seen.
This has been wrong since the introduction of physical replication, so
backpatch all the way back. In stable branches, keep the new
XLogReaderState members at the end of the struct, to avoid an ABI
break.
Author: Álvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
Reviewed-by: Nathan Bossart <bossartn@amazon.com>
Discussion: https://postgr.es/m/202108232252.dh7uxf6oxwcy@alvherre.pgsql
That allows to include just visibilitymapdefs.h from file.c, and in turn,
remove include of postgres.h from relcache.h.
Reported-by: Andres Freund
Discussion: https://postgr.es/m/20210913232614.czafiubr435l6egi%40alap3.anarazel.de
Author: Alexander Korotkov
Reviewed-by: Andres Freund, Tom Lane, Alvaro Herrera
Backpatch-through: 13
Updates/Deletes on a partition were allowed even without replica identity
after the parent table was added to a publication. This would later lead
to an error on subscribers. The reason was that we were not invalidating
the partition's relcache and the publication information for partitions
was not getting rebuilt. Similarly, we were not invalidating the
partitions' relcache after dropping a partitioned table from a publication
which will prohibit Updates/Deletes on its partition without replica
identity even without any publication.
Reported-by: Haiying Tang
Author: Hou Zhijie and Vignesh C
Reviewed-by: Vignesh C and Amit Kapila
Backpatch-through: 13
Discussion: https://postgr.es/m/OS0PR01MB6113D77F583C922F1CEAA1C3FBD29@OS0PR01MB6113.jpnprd01.prod.outlook.com
It is not appropriate for deduplication to apply single value strategy
when triggered by a bottom-up index deletion pass. This wastes cycles
because later bottom-up deletion passes will overinterpret older
duplicate tuples that deduplication actually just skipped over "by
design". It also makes bottom-up deletion much less effective for low
cardinality indexes that happen to cross a meaningless "index has single
key value per leaf page" threshold.
To fix, slightly narrow the conditions under which deduplication's
single value strategy is considered. We already avoided the strategy
for a unique index, since our high level goal must just be to buy time
for VACUUM to run (not to buy space). We'll now also avoid it when we
just had a bottom-up pass that reported failure. The two cases share
the same high level goal, and already overlapped significantly, so this
approach is quite natural.
Oversight in commit d168b666, which added bottom-up index deletion.
Author: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/CAH2-WznaOvM+Gyj-JQ0X=JxoMDxctDTYjiEuETdAGbF5EUc3MA@mail.gmail.com
Backpatch: 14-, where bottom-up deletion was introduced.
I noticed that commit 0bead9af48 left this flag undocumented in
XLogSetRecordFlags, which led me to discover that the flag doesn't
actually do what the one comment on it said it does. Improve the
situation by adding some more comments.
Backpatch to 14, where the aforementioned commit appears.
Author: Álvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/202109212119.c3nhfp64t2ql@alvherre.pgsql
Session statistics, as introduced by 960869da08, had several shortcomings:
- an additional GetCurrentTimestamp() call that also impaired the accuracy of
the data collected
This can be avoided by passing the current timestamp we already have in
pgstat_report_stat().
- an additional statistics UDP packet sent every 500ms
This is solved by adding the new statistics to PgStat_MsgTabstat.
This is conceptually ugly, because session statistics are not
table statistics. But the struct already contains data unrelated
to tables, so there is not much damage done.
Connection and disconnection are reported in separate messages, which
reduces the number of additional messages to two messages per session and a
slight increase in PgStat_MsgTabstat size (but the same number of table
stats fit).
- Session time computation could overflow on systems where long is 32 bit.
Reported-By: Andres Freund <andres@anarazel.de>
Author: Andres Freund <andres@anarazel.de>
Author: Laurenz Albe <laurenz.albe@cybertec.at>
Discussion: https://postgr.es/m/20210801205501.nyxzxoelqoo4x2qc%40alap3.anarazel.de
Backpatch: 14-, where the feature was introduced.
Formerly, we sent signals for outgoing NOTIFY messages within
ProcessCompletedNotifies, which was also responsible for sending
relevant ones of those messages to our connected client. It therefore
had to run during the main-loop processing that occurs just before
going idle. This arrangement had two big disadvantages:
* Now that procedures allow intra-command COMMITs, it would be
useful to send NOTIFYs to other sessions immediately at COMMIT
(though, for reasons of wire-protocol stability, we still shouldn't
forward them to our client until end of command).
* Background processes such as replication workers would not send
NOTIFYs at all, since they never execute the client communication
loop. We've had requests to allow triggers running in replication
workers to send NOTIFYs, so that's a problem.
To fix these things, move transmission of outgoing NOTIFY signals
into AtCommit_Notify, where it will happen during CommitTransaction.
Also move the possible call of asyncQueueAdvanceTail there, to
ensure we don't bloat the async SLRU if a background worker sends
many NOTIFYs with no one listening.
We can also drop the call of asyncQueueReadAllNotifications,
allowing ProcessCompletedNotifies to go away entirely. That's
because commit 790026972 added a call of ProcessNotifyInterrupt
adjacent to PostgresMain's call of ProcessCompletedNotifies,
and that does its own call of asyncQueueReadAllNotifications,
meaning that we were uselessly doing two such calls (inside two
separate transactions) whenever inbound notify signals coincided
with an outbound notify. We need only set notifyInterruptPending
to ensure that ProcessNotifyInterrupt runs, and we're done.
The existing documentation suggests that custom background workers
should call ProcessCompletedNotifies if they want to send NOTIFY
messages. To avoid an ABI break in the back branches, reduce it
to an empty routine rather than removing it entirely. Removal
will occur in v15.
Although the problems mentioned above have existed for awhile,
I don't feel comfortable back-patching this any further than v13.
There was quite a bit of churn in adjacent code between 12 and 13.
At minimum we'd have to also backpatch 51004c717, and a good deal
of other adjustment would also be needed, so the benefit-to-risk
ratio doesn't look attractive.
Per bug #15293 from Michael Powers (and similar gripes from others).
Artur Zakirov and Tom Lane
Discussion: https://postgr.es/m/153243441449.1404.2274116228506175596@wrigleys.postgresql.org
It's possible for us to copy an AlternativeSubPlan expression node
into multiple places, for example the scan quals of several
partition children. Then it's possible that we choose a different
one of the alternatives as optimal in each place. Commit 41efb8340
failed to consider this scenario, so its attempt to remove "unused"
subplans could remove subplans that were still used elsewhere.
Fix by delaying the removal logic until we've examined all the
AlternativeSubPlans in a given query level. (This does assume that
AlternativeSubPlans couldn't get copied to other query levels, but
for the foreseeable future that's fine; cf qual_is_pushdown_safe.)
Per report from Rajkumar Raghuwanshi. Back-patch to v14
where the faulty logic came in.
Discussion: https://postgr.es/m/CAKcux6==O3NNZC3bZ2prRYv3cjm3_Zw1GfzmOjEVqYN4jub2+Q@mail.gmail.com
If an allocation failed within LLVM it is not safe to call back into LLVM as
LLVM is not generally safe against exceptions / stack-unwinding. Thus errors
while in LLVM code are promoted to FATAL. However llvm_shutdown() did call
back into LLVM even in such cases, while llvm_release_context() was careful
not to do so.
We cannot generally skip shutting down LLVM, as that can break profiling. But
it's OK to do so if there was an error from within LLVM.
Reported-By: Jelte Fennema <Jelte.Fennema@microsoft.com>
Author: Andres Freund <andres@anarazel.de>
Author: Justin Pryzby <pryzby@telsasoft.com>
Discussion: https://postgr.es/m/AM5PR83MB0178C52CCA0A8DEA0207DC14F7FF9@AM5PR83MB0178.EURPRD83.prod.outlook.com
Backpatch: 11-, where jit was introduced
Commit 01e658fa74 added hash support for row types. This also added
support for hashing anonymous record types, using the same approach
that the type cache uses for comparison support for record types: It
just reports that it works, but it might fail at run time if a
component type doesn't actually support the operation. We get away
with that for comparison because most types support that. But some
types don't support hashing, so the current state can result in
failures at run time where the planner chooses hashing over sorting,
whereas that previously worked if only sorting was an option.
We do, however, want the record hashing support for path tracking in
recursive unions, and the SEARCH and CYCLE clauses built on that. In
that case, hashing is the only plan option. So enable that, this
commit implements the following approach: The type cache does not
report that hashing is available for the record type. This undoes
that part of 01e658fa74. Instead, callers that require hashing no
matter what can override that result themselves. This patch only
touches the callers to make the aforementioned recursive query cases
work, namely the parse analysis of unions, as well as the hash_array()
function.
Reported-by: Sait Talha Nisanci <sait.nisanci@microsoft.com>
Bug: #17158
Discussion: https://www.postgresql.org/message-id/flat/17158-8a2ba823982537a4%40postgresql.org
Updates/Deletes on a relation were allowed even without replica identity
after we define the publication for all tables. This would later lead to
an error on subscribers. The reason was that for such publications we were
not invalidating the relcache and the publication information for
relations was not getting rebuilt. Similarly, we were not invalidating the
relcache after dropping of such publications which will prohibit
Updates/Deletes without replica identity even without any publication.
Author: Vignesh C and Hou Zhijie
Reviewed-by: Hou Zhijie, Kyotaro Horiguchi, Amit Kapila
Backpatch-through: 10, where it was introduced
Discussion: https://postgr.es/m/CALDaNm0pF6zeWqCA8TCe2sDuwFAy8fCqba=nHampCKag-qLixg@mail.gmail.com
543f36b43d was too hasty in thinking that the volatility of date_bin()
had to match date_trunc(), since only the latter references
session_timezone.
Bump catversion
Per feedback from Aleksander Alekseev
Backpatch to v14, as the former commit was
