An array element equal to INT_MAX gave this code indigestion,
causing an infinite loop that surely ended in SIGSEGV. We fixed
some nearby problems awhile ago (cf 757c5182f) but missed this.
Report and diagnosis by Alexander Lakhin (bug #18273); patch by me
Discussion: https://postgr.es/m/18273-9a832d1da122600c@postgresql.org
During the calculations of the maximum for the number of buckets, take into
account that later we round that to the next power of 2.
Reported-by: Karen Talarico
Bug: #16925
Discussion: https://postgr.es/m/16925-ec96d83529d0d629%40postgresql.org
Author: Thomas Munro, Andrei Lepikhov, Alexander Korotkov
Reviewed-by: Alena Rybakina
Backpatch-through: 12
CheckPWChallengeAuth() would return STATUS_ERROR if the user does not
exist or has no password assigned, even if the client disconnected
without responding to the password challenge (as libpq often will,
for example). We should return STATUS_EOF in that case, and the
lower-level functions do, but this code level got it wrong since the
refactoring done in 7ac955b34. This breaks the intent of not logging
anything for EOF cases (cf. comments in auth_failed()) and might
also confuse users of ClientAuthentication_hook.
Per report from Liu Lang. Back-patch to all supported versions.
Discussion: https://postgr.es/m/b725238c-539d-cb09-2bff-b5e6cb2c069c@esgyn.cn
If the underlying table isn't being dumped, it's useless to dump
an extended statistics object; it'll just cause errors at restore.
We have always applied similar policies to, say, indexes.
(When and if we get cross-table stats objects, it might be profitable
to think a little harder about what to do with them. But for now
there seems no point in considering a stats object as anything but
an appendage of its table.)
Rian McGuire and Tom Lane, per report from Rian McGuire.
Back-patch to supported branches.
Discussion: https://postgr.es/m/7075d3aa-3f05-44a5-b68f-47dc6a8a0550@buildkite.com
The documentation has been missing one value in the list of catalog OIDs
that can be given to the validator function of a FDW, as of
AttributeRelationId, when changing the attribute options of a foreign
table.
Author: Ian Lawrence Barwick
Discussion: https://postgr.es/m/CAB8KJ=i16t2yJU_Pq2Z+hnNGWFhagp_bJmzxHZu3ZkOjZm-+rQ@mail.gmail.com
Backpatch-through: 12
set_config_option() bails out early if it detects that the option to
be set is PGC_BACKEND or PGC_SU_BACKEND class and we're reading the
config file in a postmaster child; we don't want to apply any new
value in such a case. That's fine as far as it goes, but it fails
to consider the requirements of the pg_file_settings view: for that,
we need to check validity of the value even though we have no
intention to apply it. Because we didn't, even very silly values
for affected GUCs would be reported as valid by the view. There
are only half a dozen such GUCs, which perhaps explains why this
got overlooked for so long.
Fix by continuing when changeVal is false; this parallels the logic
in some other early-exit paths.
Also, the check added by commit 924bcf4f1 to prevent GUC changes in
parallel workers seems a few bricks shy of a load: it's evidently
assuming that ereport(elevel, ...) won't return. Make sure we
bail out if it does. The lack of trouble reports suggests that
this is only a latent bug, i.e. parallel workers don't actually
reach here with elevel < ERROR. (Per the code coverage report,
we never reach here at all in the regression suite.) But we clearly
don't want to risk proceeding if that does happen.
Per report from Rıdvan Korkmaz. These are ancient bugs, so back-patch
to all supported branches.
Discussion: https://postgr.es/m/2089235.1703617353@sss.pgh.pa.us
Like commit 388e80132, use "#pragma GCC system_header" to silence
warnings appearing within the Python headers, since newer Python
versions no longer worry about some restrictions we still use like
-Wdeclaration-after-statement.
This patch improves on 388e80132 by inventing a separate wrapper
header file, allowing the pragma to be tightly scoped to just
the Python headers and not other stuff we have laying about in
plpython.h. I applied the same technique to plperl for the same
reason: the original patch suppressed warnings for a good deal
of our own code, not only the Perl headers.
Like the previous commit, back-patch to supported branches.
Peter Eisentraut and Tom Lane
Discussion: https://postgr.es/m/ae523163-6d2a-4b81-a875-832e48dec502@eisentraut.org
This is necessary when spgcanreturn() is invoked on a partitioned
index, and the failure might be reachable in other scenarios as
well. The rest of what spgGetCache() does is perfectly sensible
for a partitioned index, so we should allow it to go through.
I think the main takeaway from this is that we lack sufficient test
coverage for non-btree partitioned indexes. Therefore, I added
simple test cases for brin and gin as well as spgist (hash and
gist AMs were covered already in indexing.sql).
Per bug #18256 from Alexander Lakhin. Although the known test case
only fails since v16 (3c569049b), I've got no faith at all that there
aren't other ways to reach this problem; so back-patch to all
supported branches.
Discussion: https://postgr.es/m/18256-0b0e1b6e4a620f1b@postgresql.org
The example for dropping an option was incorrectly quoting the
option key thus making it a value turning the command into an
unqualified ADD operation. The result of dropping became adding
a new key/value pair instead:
d=# alter foreign data wrapper f options (drop 'b');
ALTER FOREIGN DATA WRAPPER
d=# select fdwoptions from pg_foreign_data_wrapper where fdwname='f';
fdwoptions
------------
{drop=b}
(1 row)
This has been incorrect for a long time so backpatch to all
supported branches.
Author: Tim <tim.needham2@gmail.com>
Discussion: https://postgr.es/m/170292280173.1876505.5204623074024041738@wrigleys.postgresql.org
This function reads directly a page from a relation, relying on
index_open() to open the index to read from. Unfortunately, this would
crash when using partitioned indexes, as these can be opened with
index_open() but they have no physical pages.
Alexander has fixed the module, while I have written the test.
Author: Alexander Lakhin, Michael Paquier
Discussion: https://postgr.es/m/18246-f4d9ff7cb3af77e6@postgresql.org
Backpatch-through: 12
As coded, the function relied on index_open() when opening an index
relation, allowing partitioned indexes to be processed by
pgstathashindex(). This was leading to a "could not open file" error
because partitioned indexes have no physical files, or to a crash with
an assertion failure (like on HEAD).
This issue is fixed by applying the same checks as the other stat
functions for indexes, with a lookup at both RELKIND_INDEX and the index
AM expected.
Author: Alexander Lakhin
Discussion: https://postgr.es/m/18246-f4d9ff7cb3af77e6@postgresql.org
Backpatch-through: 12
We didn't explain this clearly until somewhere deep in the
"Extending SQL" chapter, but really it ought to be mentioned
in the introductory material too.
Discussion: https://postgr.es/m/4097442.1694967650@sss.pgh.pa.us
In v16 and up (since commit afbfc0298), large object ownership
checking has been broken because object_ownercheck() didn't take care
of the discrepancy between our object-address representation of large
objects (classId == LargeObjectRelationId) and the catalog where their
ownership info is actually stored (LargeObjectMetadataRelationId).
This resulted in failures such as "unrecognized class ID: 2613"
when trying to update blob properties as a non-superuser.
Poking around for related bugs, I found that AlterObjectOwner_internal
would pass the wrong classId to the PostAlterHook in the no-op code
path where the large object already has the desired owner. Also,
recordExtObjInitPriv checked for the wrong classId; that bug is only
latent because the stanza is dead code anyway, but as long as we're
carrying it around it should be less wrong. These bugs are quite old.
In HEAD, we can reduce the scope for future bugs of this ilk by
changing AlterObjectOwner_internal's API to let the translation happen
inside that function, rather than requiring callers to know about it.
A more bulletproof fix, perhaps, would be to start using
LargeObjectMetadataRelationId as the dependency and object-address
classId for blobs. However that has substantial risk of breaking
third-party code; even within our own code, it'd create hassles
for pg_dump which would have to cope with a version-dependent
representation. For now, keep the status quo.
Discussion: https://postgr.es/m/2650449.1702497209@sss.pgh.pa.us
Dead tuples are ignored and are not marked as dead during recovery, as
it can lead to MVCC issues on a standby because its xmin may not match
with the primary. This information is tracked by a field called
"xactStartedInRecovery" in the transaction state data, switched on when
starting a transaction in recovery.
Unfortunately, this information was not correctly tracked when starting
a subtransaction, because the transaction state used for the
subtransaction did not update "xactStartedInRecovery" based on the state
of its parent. This would cause index scans done in subtransactions to
return inconsistent data, depending on how the xmin of the primary
and/or the standby evolved.
This is broken since the introduction of hot standby in efc16ea520, so
backpatch all the way down.
Author: Fei Changhong
Reviewed-by: Kyotaro Horiguchi
Discussion: https://postgr.es/m/tencent_C4D907A5093C071A029712E73B43C6512706@qq.com
Backpatch-through: 12
OpenSSL will sometimes return SSL_ERROR_SYSCALL without having set
errno; this is apparently a reflection of recv(2)'s habit of not
setting errno when reporting EOF. Ensure that we treat such cases
the same as read EOF. Previously, we'd frequently report them like
"could not accept SSL connection: Success" which is confusing, or
worse report them with an unrelated errno left over from some
previous syscall.
To fix, ensure that errno is zeroed immediately before the call,
and report its value only when it's not zero afterwards; otherwise
report EOF.
For consistency, I've applied the same coding pattern in libpq's
pqsecure_raw_read(). Bare recv(2) shouldn't really return -1 without
setting errno, but in case it does we might as well cope.
Per report from Andres Freund. Back-patch to all supported versions.
Discussion: https://postgr.es/m/20231208181451.deqnflwxqoehhxpe@awork3.anarazel.de
The apply worker needs to update the state of the subscription tables to
'READY' during the synchronization phase which requires locking the
corresponding subscription. The apply worker also waits for the
subscription tables to reach the 'SYNCDONE' state after holding the locks
on the subscription and the wait is done using WaitLatch. The 'SYNCDONE'
state is changed by tablesync workers again by locking the corresponding
subscription. Both the state updates use AccessShareLock mode to lock the
subscription, so they can't block each other. However, a backend can
simultaneously try to acquire a lock on the same subscription using
AccessExclusiveLock mode to alter the subscription. Now, the backend's
wait on a lock can sneak in between the apply worker and table sync worker
causing deadlock.
In other words, apply_worker waits for tablesync worker which waits for
backend, and backend waits for apply worker. This is not detected by the
deadlock detector because apply worker uses WaitLatch.
The fix is to release existing locks in apply worker before it starts to
wait for tablesync worker to change the state.
Reported-by: Tomas Vondra
Author: Shlok Kyal
Reviewed-by: Amit Kapila, Peter Smith
Backpatch-through: 12
Discussion: https://postgr.es/m/d291bb50-12c4-e8af-2af2-7bb9bb4d8e3e@enterprisedb.com
Commit 5a991ef869 accidentally reversed the order of the tuples
and fields parameters, making the error message incorrectly refer
to 3 tuples with 1 field when IDENTIFY_SYSTEM returns 1 tuple and
3 or 4 fields. Fix by changing the order of the parameters. This
also adds a comment describing why we check for < 3 when postgres
since 9.4 has been sending 4 fields.
Backpatch all the way since the bug is almost a decade old.
Author: Tomonari Katsumata <t.katsumata1122@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Bug: #18224
Backpatch-through: v12
The wording changed here comes from 991bfe11d2, when the only way to
trigger a promotion was with a trigger file. There are more options to
achieve this operation these days, like the SQL function pg_promote() or
the command `pg_ctl promote`, so it is confusing to assume that only a
trigger file is able to do the work.
Note also that promote_trigger_file has been removed as of cd4329d939
in 16~.
Author: Shinya Kato
Discussion: https://postgr.es/m/201b08ea29aa61f96162080e75be503c@oss.nttdata.com
Backpatch-through: 12
When creating a partitioned index, the partition key must be a subset
of the index's columns. But this currently doesn't check that the
collations between the partition key and the index definition match.
So you can construct a unique index that fails to enforce uniqueness.
(This would most likely involve a nondeterministic collation, so it
would have to be crafted explicitly and is not something that would
just happen by accident.)
This patch adds the required collation check. As a result, any
previously allowed unique index that has a collation mismatch would no
longer be allowed to be created.
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/flat/3327cb54-f7f1-413b-8fdb-7a9dceebb938%40eisentraut.org
We should have done it this way all along, but we accidentally got
away with using the wrong BIO field up until OpenSSL 3.2. There,
the library's BIO routines that we rely on use the "data" field
for their own purposes, and our conflicting use causes assorted
weird behaviors up to and including core dumps when SSL connections
are attempted. Switch to using the approved field for the purpose,
i.e. app_data.
While at it, remove our configure probes for BIO_get_data as well
as the fallback implementation. BIO_{get,set}_app_data have been
there since long before any OpenSSL version that we still support,
even in the back branches.
Also, update src/test/ssl/t/001_ssltests.pl to allow for a minor
change in an error message spelling that evidently came in with 3.2.
Tristan Partin and Bo Andreson. Back-patch to all supported branches.
Discussion: https://postgr.es/m/CAN55FZ1eDDYsYaL7mv+oSLUij2h_u6hvD4Qmv-7PK7jkji0uyQ@mail.gmail.com
If the tuple being updated is not visible to the crosscheck snapshot,
we return TM_Updated but the assertions would not hold in that case.
Move them to before the cross-check.
Fixes bug #17893. Backpatch to all supported versions.
Author: Alexander Lakhin
Backpatch-through: 12
Discussion: https://www.postgresql.org/message-id/17893-35847009eec517b5%40postgresql.org
The libpq code in charge of creating per-connection SSL objects was
prone to a race condition when loading the custom BIO methods needed by
my_SSL_set_fd(). As BIO methods are stored as a static variable, the
initialization of a connection could fail because it could be possible
to have one thread refer to my_bio_methods while it is being manipulated
by a second concurrent thread.
This error has been introduced by 8bb14cdd33, that has removed
ssl_config_mutex around the call of my_SSL_set_fd(), that itself sets
the custom BIO methods used in libpq. Like previously, the BIO method
initialization is now protected by the existing ssl_config_mutex, itself
initialized earlier for WIN32.
While on it, document that my_bio_methods is protected by
ssl_config_mutex, as this can be easy to miss.
Reported-by: Willi Mann
Author: Willi Mann, Michael Paquier
Discussion: https://postgr.es/m/e77abc4c-4d03-4058-a9d7-ef0035657e04@celonis.com
Backpatch-through: 12
When using GSSAPI encryption in non-blocking mode, libpq sometimes
failed with "GSSAPI caller failed to retransmit all data needing
to be retried". The cause is that pqPutMsgEnd rounds its transmit
request down to an even multiple of 8K, and sometimes that can lead
to not requesting a write of data that was requested to be written
(but reported as not written) earlier. That can upset pg_GSS_write's
logic for dealing with not-yet-written data, since it's possible
the data in question had already been incorporated into an encrypted
packet that we weren't able to send during the previous call.
We could fix this with a one-or-two-line hack to disable pqPutMsgEnd's
round-down behavior, but that seems like making the caller work around
a behavior that pg_GSS_write shouldn't expose in this way. Instead,
adjust pg_GSS_write to never report a partial write: it either
reports a complete write, or reflects the failure of the lower-level
pqsecure_raw_write call. The requirement still exists for the caller
to present at least as much data as on the previous call, but with
the caller-visible write start point not moving there is no temptation
for it to present less. We lose some ability to reclaim buffer space
early, but I doubt that that will make much difference in practice.
This also gets rid of a rather dubious assumption that "any
interesting failure condition (from pqsecure_raw_write) will recur
on the next try". We've not seen failure reports traceable to that,
but I've never trusted it particularly and am glad to remove it.
Make the same adjustments to the equivalent backend routine
be_gssapi_write(). It is probable that there's no bug on the backend
side, since we don't have a notion of nonblock mode there; but we
should keep the logic the same to ease future maintenance.
Per bug #18210 from Lars Kanis. Back-patch to all supported branches.
Discussion: https://postgr.es/m/18210-4c6d0b14627f2eb8@postgresql.org
As written, the query checked for an access method of type 's', which is
not an AM type supported in the core code.
Error introduced by 8586bf7ed8. As this query is not checking what it
should, backpatch all the way down.
Reviewed-by: Aleksander Alekseev
Discussion: https://postgr.es/m/ZVxJkAJrKbfHETiy@paquier.xyz
Backpatch-through: 12
The DROP STATISTICS code failed to properly lock the table, leading to
ERROR: tuple concurrently deleted
when executed concurrently with ANALYZE.
Fixed by modifying RemoveStatisticsById() to acquire the same lock as
ANALYZE. This function is called only by DROP STATISTICS, as ANALYZE
calls RemoveStatisticsDataById() directly.
Reported by Justin Pryzby, fix by me. Backpatch through 12. The code was
like this since it was introduced in 10, but older releases are EOL.
Reported-by: Justin Pryzby
Reviewed-by: Tom Lane
Backpatch-through: 12
Discussion: https://postgr.es/m/ZUuk-8CfbYeq6g_u@pryzbyj2023
Commits 146604ec43 and a898b409f6 added overflow checks to
interval_mul(), but not to interval_div(), which contains almost
identical code, and so is susceptible to the same kinds of
overflows. In addition, those checks did not catch all possible
overflow conditions.
Add additional checks to the "cascade down" code in interval_mul(),
and copy all the overflow checks over to the corresponding code in
interval_div(), so that they both generate "interval out of range"
errors, rather than returning bogus results.
Given that these errors are relatively easy to hit, back-patch to all
supported branches.
Per bug #18200 from Alexander Lakhin, and subsequent investigation.
Discussion: https://postgr.es/m/18200-5ea288c7b2d504b1%40postgresql.org
When performing inlining LLVM unfortunately "leaks" types (the
types survive and are usable, but a new round of inlining will
recreate new structurally equivalent types). This accumulation
will over time amount to a memory leak which for some queries
can be large enough to trigger the OOM process killer.
To avoid accumulation of types, all IR related data is stored
in an LLVMContextRef which is dropped and recreated in order
to release all types. Dropping and recreating incurs overhead,
so it will be done only after 100 queries. This is a heuristic
which might be revisited, but until we can get the size of the
context from LLVM we are flying a bit blind.
This issue has been reported several times, there may be more
references to it in the archives on top of the threads linked
below.
This is a backpatch of 9dce22033d to all supported branches.
Reported-By: Justin Pryzby <pryzby@telsasoft.com>
Reported-By: Kurt Roeckx <kurt@roeckx.be>
Reported-By: Jaime Casanova <jcasanov@systemguards.com.ec>
Reported-By: Lauri Laanmets <pcspets@gmail.com>
Author: Andres Freund and Daniel Gustafsson
Discussion: https://postgr.es/m/7acc8678-df5f-4923-9cf6-e843131ae89d@www.fastmail.com
Discussion: https://postgr.es/m/20201218235607.GC30237@telsasoft.com
Discussion: https://postgr.es/m/CAPH-tTxLf44s3CvUUtQpkDr1D8Hxqc2NGDzGXS1ODsfiJ6WSqA@mail.gmail.com
Backpatch-through: v12
This seems more correct, because other before_shmem_exit calls may
expect the infrastructure that is needed to run queries and access the
database to be working, and also because this cleanup has nothing to
do with shared memory.
This is a back-patch of bab150045b.
There were no known user-visible consequences to this, though, apart
from what was previous fixed by commit 303640199d and back-patched
as commit bcbc27251d and commit f7013683d9, so bab150045b was
not no back-patched at the time.
Bharath Rupireddy
Discussion: http://postgr.es/m/CALj2ACWk7j4F2v2fxxYfrroOF=AdFNPr1WsV+AGtHAFQOqm_pw@mail.gmail.com
Backpatch-through: 13, 12
contain_mutable_functions and contain_volatile_functions give
reliable answers only after expression preprocessing (specifically
eval_const_expressions). Some places understand this, but some did
not get the memo --- which is not entirely their fault, because the
problem is documented only in places far away from those functions.
Introduce wrapper functions that allow doing the right thing easily,
and add commentary in hopes of preventing future mistakes from
copy-and-paste of code that's only conditionally safe.
Two actual bugs of this ilk are fixed here. We failed to preprocess
column GENERATED expressions before checking mutability, so that the
code could fail to detect the use of a volatile function
default-argument expression, or it could reject a polymorphic function
that is actually immutable on the datatype of interest. Likewise,
column DEFAULT expressions weren't preprocessed before determining if
it's safe to apply the attmissingval mechanism. A false negative
would just result in an unnecessary table rewrite, but a false
positive could allow the attmissingval mechanism to be used in a case
where it should not be, resulting in unexpected initial values in a
new column.
In passing, re-order the steps in ComputePartitionAttrs so that its
checks for invalid column references are done before applying
expression_planner, rather than after. The previous coding would
not complain if a partition expression contains a disallowed column
reference that gets optimized away by constant folding, which seems
to me to be a behavior we do not want.
Per bug #18097 from Jim Keener. Back-patch to all supported versions.
Discussion: https://postgr.es/m/18097-ebb179674f22932f@postgresql.org
The fallback implementation of pg_atomic_test_set_flag() that uses
atomic-exchange gives pg_atomic_exchange_u32_impl() an extra
argument. This issue has been present since the introduction of
the atomics API in commit b64d92f1a5.
Reviewed-by: Andres Freund
Discussion: https://postgr.es/m/20231114035439.GA1809032%40nathanxps13
Backpatch-through: 12
checkExtensionMembership() set the DUMP_COMPONENT_SECLABEL and
DUMP_COMPONENT_POLICY flags for extension member objects, even though
we lack any infrastructure for tracking extensions' initial settings
of these properties. This is not OK. The result was that a dump
would always include commands to set these properties for extension
objects that have them, with at least three negative consequences:
1. The restoring user might not have privilege to set these properties
on these objects.
2. The properties might be incorrect/irrelevant for the version of the
extension that's installed in the destination database.
3. The dump itself might fail, in the case of RLS properties attached
to extension tables that the dumping user lacks privilege to LOCK.
(That's because we must get at least AccessShareLock to ensure that
we don't fail while trying to decompile the RLS expressions.)
When and if somebody cares to invent initial-state infrastructure for
extensions' RLS policies and security labels, we could think about
finding another way around problem #3. But in the absence of such
infrastructure, this whole thing is just wrong and we shouldn't do it.
(Note: this applies only to ordinary dumps; binary-upgrade dumps
still dump and restore extension member objects separately, with
all properties.)
Tom Lane and Jacob Champion. Back-patch to all supported branches.
Discussion: https://postgr.es/m/00d46a48-3324-d9a0-49bf-e7f0f11d1038@timescale.com
It's clearly stated in the comments that ginFindParents() must keep
the pin on the index's root page that's associated with the topmost
GinBtreeStack item. However, the code path for the case that the
desired downlink has been pushed down to the next index level
ignored this proviso, and would release the pin anyway if we were
still examining the root level. That led to an assertion failure
or "buffer NNNN is not owned by resource owner" error later, when
we try to release the pin again at the end of the insertion.
This is quite hard to reproduce, since it can only happen if an
index root page split occurs concurrently with our own insertion.
Thanks to Jeff Janes for finding a test case that triggers it
often enough to allow investigation.
This has been there since the beginning of GIN, so back-patch
to all supported branches.
Discussion: https://postgr.es/m/CAMkU=1yCAKtv86dMrD__Ja-7KzjE=uMeKX8y__cx5W-OEWy2ow@mail.gmail.com
One of the examples on the SELECT page was missing a semicolon from
a listing which has the look and feel of being a psql session. This
adds the missing semicolon and also removes the newline between the
query and results to match the other examples nearby.
Backpatch to all supported branches to avoid backpatching issues on
this page.
Reported-by: tim.needham2@gmail.com
Discussion: https://postgr.es/m/169965004097.225187.12941375915673151540@wrigleys.postgresql.org
Backpatch-through: v12
Commit b7eda3e0e moved XidInMVCCSnapshot() from tqual.c into snapmgr.c,
but follow-up commit c91560def incorrectly updated this reference. We
could fix it, but as pointed out by Daniel Gustafsson, 1) the reader can
easily find the file that contains the definition of that function, e.g.
by grepping, and 2) this kind of reference is prone to going stale; so
let's just remove it.
Back-patch to all supported branches.
Reviewed by Daniel Gustafsson.
Discussion: https://postgr.es/m/CAPmGK145VdKkPBLWS2urwhgsfidbSexwY-9zCL6xSUJH%2BBTUUg%40mail.gmail.com
When computing "0 - INT64_MIN", most platforms would report an
overflow error, which is correct. However, platforms without integer
overflow builtins or 128-bit integers would fail to spot the overflow,
and incorrectly return INT64_MIN.
Back-patch to all supported branches.
Patch be me. Thanks to Jian He for initial investigation, and Laurenz
Albe and Tom Lane for review.
Discussion: https://postgr.es/m/CAEZATCUNK-AZSD0jVdgkk0N%3DNcAXBWeAEX-QU9AnJPensikmdQ%40mail.gmail.com
Tom Lane
Alexander Korotkov
Bruce Momjian
Michael Paquier
Amit Kapila
Daniel Gustafsson
Peter Eisentraut
Heikki Linnakangas
Alvaro Herrera
Tomas Vondra
Noah Misch
Dean Rasheed
Nathan Bossart
Etsuro Fujita
David Rowley