Instead of serializing the transaction to disk after reaching the
logical_decoding_work_mem limit in memory, we consume the changes we have
in memory and invoke stream API methods added by commit 45fdc973.
However, sometimes if we have incomplete toast or speculative insert we
spill to the disk because we can't generate the complete tuple and stream.
And, as soon as we get the complete tuple we stream the transaction
including the serialized changes.

We can do this incremental processing thanks to having assignments
(associating subxact with toplevel xacts) in WAL right away, and
thanks to logging the invalidation messages at each command end. These
features are added by commits 0bead9af and c55040cc respectively.

Now that we can stream in-progress transactions, the concurrent aborts
may cause failures when the output plugin consults catalogs (both system
and user-defined).

We handle such failures by returning ERRCODE_TRANSACTION_ROLLBACK
sqlerrcode from system table scan APIs to the backend or WALSender
decoding a specific uncommitted transaction. The decoding logic on the
receipt of such a sqlerrcode aborts the decoding of the current
transaction and continue with the decoding of other transactions.

We have ReorderBufferTXN pointer in each ReorderBufferChange by which we
know which xact it belongs to.  The output plugin can use this to decide
which changes to discard in case of stream_abort_cb (e.g. when a subxact
gets discarded).

We also provide a new option via SQL APIs to fetch the changes being
streamed.

Author: Dilip Kumar, Tomas Vondra, Amit Kapila, Nikhil Sontakke
Reviewed-by: Amit Kapila, Kuntal Ghosh, Ajin Cherian
Tested-by: Neha Sharma, Mahendra Singh Thalor and Ajin Cherian
Discussion: https://postgr.es/m/688b0b7f-2f6c-d827-c27b-216a8e3ea700@2ndquadrant.com

Make the nbtree page split REDO routine consistent with original
execution in its approach to acquiring and releasing buffer locks (at
least for pages on the tree level of the page being split).  This brings
btree_xlog_split() in line with btree_xlog_unlink_page(), which was
taught to couple buffer locks by commit 9a9db08a.

Note that the precise order in which we both acquire and release sibling
buffer locks in btree_xlog_split() now matches original execution
exactly (the precise order in which the locks are released probably
doesn't matter much, but we might as well be consistent about it).

The rule for nbtree REDO routines from here on is that same-level locks
should be acquired in an order that's consistent with original
execution.  It's not practical to have a similar rule for cross-level
page locks, since for the most part original execution holds those locks
for a period that spans multiple atomic actions/WAL records.  It's also
not necessary, because clearly the cross-level lock coupling is only
truly needed during original execution because of the presence of
concurrent inserters.

This is not a bug fix (unlike the similar aforementioned commit, commit
9a9db08a).  The immediate reason to tighten things up in this area is to
enable an upcoming enhancement to contrib/amcheck that allows it to
verify that sibling links are in agreement with only an AccessShareLock
(this check produced false positives when run on a replica server on
account of the inconsistency fixed by this commit).  But that's not the
only reason to be stricter here.

It is generally useful to make locking on replicas be as close to what
happens during original execution as practically possible.  It makes it
less likely that hard to catch bugs will slip in in the future.  The
previous state of affairs seems to be a holdover from before the
introduction of Hot Standby, when buffer lock acquisitions during
recovery were totally unnecessary.  See also: commit 3bbf668d, which
tightened things up in this area a few years after the introduction of
Hot Standby.

Discussion: https://postgr.es/m/CAH2-Wz=465cJj11YXD9RKH8z=nhQa2dofOZ_23h67EXUGOJ00Q@mail.gmail.com

These flags are unused and always have been.

Discussion: https://postgr.es/m/20200805235549.GA8118@alvherre.pgsql

We have various cases where we allow DDL on tables to be performed with
less than full AccessExclusiveLock.  This requires concurrent queries
to be able to cope with the DDL change mid-flight, but up to now we had
no repeatable way to test such cases.  To improve that, invent a test
module that allows halting a backend after planning and then resuming
execution once we've done desired actions in another session.  (The same
approach could be used to inject delays in other places, if there's a
suitable hook available.)

This commit includes a single test case, which is meant to exercise the
previously-untestable ExecCreatePartitionPruneState code repaired by
commit 7a980dfc.  We'd probably not bother with this if that were the
only foreseen benefit, but I expect additional test cases will use this
infrastructure in the future.

Test module by Andy Fan, partition-addition test case by me.

Discussion: https://postgr.es/m/20200802181131.GA27754@telsasoft.com

Make the nbtree page split REDO routine variable names consistent with
_bt_split() (which handles the original execution of page splits).
These names make the code easier to follow by making the distinction
between the original page and the left half of the split clear.  (The
left half of the split page is a temp page that REDO creates to replace
the origpage contents.)

Also reduce the elevel used when adding a new high key to the temp page
from PANIC to ERROR to be consistent.  We already only raise an ERROR
when data item PageAddItem() temp page calls fail.

Commit 13838740 fixed some issues with step generation in partition
pruning, but there was yet another one: get_steps_using_prefix() assumes
that clauses in the passed-in prefix list are sorted in ascending order
of their partition key numbers, but the caller failed to ensure this for
range partitioning, which led to an assertion failure in debug builds.
Adjust the caller function to arrange the clauses in the prefix list in
the required order for range partitioning.

Back-patch to v11, like the previous commit.

Patch by me, reviewed by Amit Langote.

Discussion: https://postgr.es/m/CAPmGK16jkXiFG0YqMbU66wte-oJTfW6D1HaNvQf%3D%2B5o9%3Dm55wQ%40mail.gmail.com

Oversight in commit d114cc53.

This will be helpful if it ever proves necessary to revoke an opclass's
support for deduplication.

Backpatch: 13-, where nbtree deduplication was introduced.

9bdb300d modified the EXPLAIN output for Hash Aggregate to show details
from parallel workers. However, it neglected to consider that a given
parallel worker may not have assisted with the given Hash Aggregate. This
can occur when workers fail to start or during Parallel Append with
enable_partitionwise_join enabled when only a single worker is working on
a non-parallel aware sub-plan. It could also happen if a worker simply
wasn't fast enough to get any work done before other processes went and
finished all the work.

The bogus output came from the fact that ExplainOpenWorker() skipped
showing any details for non-initialized workers but show_hashagg_info()
did show details from the worker.  This meant that the worker properties
that were shown were not properly attributed to the worker that they
belong to.

In passing, we also now don't show Hash Aggregate properties for the
leader process when it did not contribute any work to the Hash Aggregate.
This can occur either during Parallel Append when only a parallel worker
worked on a given sub plan or with parallel_leader_participation set to
off.  This aims to make the behavior of Hash Aggregate's EXPLAIN output
more similar to Sort's.

Reported-by: Justin Pryzby
Discussion: https://postgr.es/m/20200805012105.GZ28072%40telsasoft.com
Backpatch-through: 13, where the original breakage was introduced

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.

There are no known user-visible consequences to this, though, apart
from what was previous fixed by commit
30364019 and back-patched as commit
bcbc27251d35336a6442761f59638138a772b839 and commit
f7013683d9bb663a6a917421b1374306a32f165b, so for now, no back-patch.

Bharath Rupireddy

Discussion: http://postgr.es/m/CALj2ACWk7j4F2v2fxxYfrroOF=AdFNPr1WsV+AGtHAFQOqm_pw@mail.gmail.com

Reported-by: Vyacheslav Shablistyy

Discussion: https://postgr.es/m/159586122762.680.1361378513036616007@wrigleys.postgresql.org

Backpatch-through: 9.5

Since we no longer require AccessExclusiveLock to add a partition,
the executor may see that a partitioned table has more partitions
than the planner saw.  ExecCreatePartitionPruneState's code for
matching up the partition lists in such cases was faulty, and would
misbehave if the planner had successfully pruned any partitions from
the query.  (Thus, trouble would occur only if a partition addition
happens concurrently with a query that uses both static and dynamic
partition pruning.)  This led to an Assert failure in debug builds,
and probably to crashes or query misbehavior in production builds.

To repair the bug, just explicitly skip zeroes in the plan's
relid_map[] list.  I also made some cosmetic changes to make the code
more readable (IMO anyway).  Also, convert the cross-checking Assert
to a regular test-and-elog, since it's now apparent that this logic
is more fragile than one would like.

Currently, there's no way to repeatably exercise this code, except
with manual use of a debugger to stop the backend between planning
and execution.  Hence, no test case in this patch.  We oughta do
something about that testability gap, but that's for another day.

Amit Langote and Tom Lane, per report from Justin Pryzby.  Oversight
in commit 898e5e32; backpatch to v12 where that appeared.

Discussion: https://postgr.es/m/20200802181131.GA27754@telsasoft.com

Currently, page unlink leaves remaining items "as is", but replay of
corresponding WAL-record re-initializes page leaving it with no items.
For the sake of consistency, this commit makes primary delete all the items
during page unlink as well.

Thanks to this change, we now don't mask contents of deleted btree page for
WAL consistency checking.

Discussion: https://postgr.es/m/CAPpHfdt_OTyQpXaPJcWzV2N-LNeNJseNB-K_A66qG%3DL518VTFw%40mail.gmail.com
Author: Alexander Korotkov
Reviewed-by: Peter Geoghegan

A couple of test cases had connect_timeout=14, a value that seems
to have been plucked from a hat.  While it's more than sufficient
for normal cases, slow/overloaded buildfarm machines can get a timeout
failure here, as per recent report from "sungazer".  Increase to 180
seconds, which is in line with our typical timeouts elsewhere in
the regression tests.

Back-patch to 9.6; the code looks different in 9.5, and this doesn't
seem to be quite worth the effort to adapt to that.

Report: https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=sungazer&dt=2020-08-04%2007%3A12%3A22

The test would fail in an environment including a certificate file in
~/.postgresql/.  bdd6e9ba fixed a similar failure, and d6e612f8 introduced
the same problem again with a new test.

Author: Kyotaro Horiguchi
Discussion: https://postgr.es/m/20200804.120033.31225582282178001.horikyota.ntt@gmail.com
Backpatch-through: 13

It was possible for the logic used by backward scans (which must reason
about concurrent page splits/deletions in its own peculiar way) to
become confused when running on a replica.  Concurrent replay of a WAL
record that describes the second phase of page deletion could cause
_bt_walk_left() to get confused.  btree_xlog_unlink_page() simply failed
to adhere to the same locking protocol that we use on the primary, which
is obviously wrong once you consider these two disparate functions
together.  This bug is present in all stable branches.

More concretely, the problem was that nothing stopped _bt_walk_left()
from observing inconsistencies between the deletion's target page and
its original sibling pages when running on a replica.  This is true even
though the second phase of page deletion is supposed to work as a single
atomic action.  Queries running on replicas raised "could not find left
sibling of block %u in index %s" can't-happen errors when they went back
to their scan's "original" page and observed that the page has not been
marked deleted (even though it really was concurrently deleted).

There is no evidence that this actually happened in the real world.  The
issue came to light during unrelated feature development work.  Note
that _bt_walk_left() is the only code that cares about the difference
between a half-dead page and a fully deleted page that isn't also
exclusively used by nbtree VACUUM (unless you include contrib/amcheck
code).  It seems very likely that backward scans are the only thing that
could become confused by the inconsistency.  Even amcheck's complex
bt_right_page_check_scankey() dance was unaffected.

To fix, teach btree_xlog_unlink_page() to lock the left sibling, target,
and right sibling pages in that order before releasing any locks (just
like _bt_unlink_halfdead_page()).  This is the simplest possible
approach.  There doesn't seem to be any opportunity to be more clever
about lock acquisition in the REDO routine, and it hardly seems worth
the trouble in any case.

This fix might enable contrib/amcheck verification of leaf page sibling
links with only an AccessShareLock on the relation.  An amcheck patch
from Andrey Borodin was rejected back in January because it clashed with
btree_xlog_unlink_page()'s lax approach to locking pages.  It now seems
likely that the real problem was with btree_xlog_unlink_page(), not the
patch.

This is a low severity, low likelihood bug, so no backpatch.

Author: Michail Nikolaev
Diagnosed-By: Michail Nikolaev
Discussion: https://postgr.es/m/CANtu0ohkR-evAWbpzJu54V8eCOtqjJyYp3PQ_SGoBTRGXWhWRw@mail.gmail.com

Add a documenting assertion that's similar to the nearby assertion added
by commit cd8c73a3.  This conveys that the entire call to _bt_pagedel()
does no work if it isn't possible to get a descent stack for the initial
scanblkno page.

A moment's examination of these queries is sufficient to see that
they do not produce duplicate rows, unless perhaps there's
catalog corruption.  Using DISTINCT anyway is inefficient and
confusing; moreover it sets a poor example for anyone who
refers to psql -E output to see how to query the catalogs.

We've allowed UTC offsets up to +/- 15:59 since commit cd0ff9c0, but
that commit forgot to fix the documentation about timetz.

Per bug #16571 from osdba.

Discussion: https://postgr.es/m/16571-eb7501598de78c8a@postgresql.org

This ought to work much like C's "#elif defined(name)"; but the code
implemented it in a way equivalent to endif followed by ifdef, so that
it didn't matter whether any previous branch of the IF construct had
succeeded.  Fix that; add some test cases covering elif and nested IFs;
and improve the documentation, which also seemed a bit confused.

AFAICS the code has been like this since the feature was added in 1999
(commit b57b0e04).  So while it's surely wrong, there might be code
out there relying on the current behavior.  Hence, don't back-patch
into stable branches.  It seems all right to fix it in v13 though.

Per report from Ashutosh Sharma.  Reviewed by Ashutosh Sharma and
Michael Meskes.

Discussion: https://postgr.es/m/CAE9k0P=dQk9X0cU2tN49S7a9tv733-e1pVdpB1P-pWJ5PdTktg@mail.gmail.com

This is useful for monitoring purposes with log parsing.  Similarly to
pg_stat_activity, the leader's PID is shown only for active parallel
workers, minimizing the log footprint for the leaders as the equivalent
shared memory field is set as long as a backend is alive.

Author: Justin Pryzby
Reviewed-by: Álvaro Herrera, Michael Paquier, Julien Rouhaud, Tom Lane
Discussion: https://postgr.es/m/20200315111831.GA21492@telsasoft.com

Instead of writing a query to psql's stdin, use -c.  This avoids a
failure where psql exits before we write, seen a few times on the build
farm.  Thanks to Tom Lane for the suggestion.

Back-patch to 11, where the LDAP tests arrived.
Reviewed-by: Noah Misch <noah@leadboat.com>
Discussion: https://postgr.es/m/CA%2BhUKGLFmW%2BHQYPeKiwSp5sdFFHtFViCpw4Mh6yAgEx74r5-Cw%40mail.gmail.com

Post-commit review for commit 84c0e4b9.

Author: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CAApHDvptBx_%2BUPAzY0uXzopbvPVGKPeZ6Hoy8rnPcWz20Cr0Bw%40mail.gmail.com

The type-name pattern in \dAc and \dAf was matched only to the actual
pg_type.typname string, which is fairly user-unfriendly in cases where
that is not what's shown to the user by format_type (compare "_int4"
and "integer[]").  Make this code match what \dT does, i.e. match the
pattern against either typname or format_type() output.  Also fix its
broken handling of schema-name restrictions.  (IOW, make these
processSQLNamePattern calls match \dT's.)  While here, adjust
whitespace to make the query a little prettier in -E output, too.

Also improve some inaccuracies and shaky grammar in the related
documentation.

Noted while working on a patch for intarray's opclasses; I wondered
why I couldn't get a match to "integer*" for the input type name.

Windows 64bit has 4-byte long values which is not suitable for tracking
disk space usage in the incremental sort code. Let's just make all these
fields int64s.

Author: James Coleman
Discussion: https://postgr.es/m/CAApHDvpky%2BUhof8mryPf5i%3D6e6fib2dxHqBrhp0Qhu0NeBhLJw%40mail.gmail.com
Backpatch-through: 13, where the incremental sort code was added

We have edge-case bugs when assigning values in the last few dozen pages
before the wrap limit.  We may introduce similar bugs in the future.  At
default BLCKSZ, this makes such bugs unreachable outside of single-user
mode.  Also, when VACUUM began to consume mxacts, multiStopLimit did not
change to compensate.

pg_upgrade may fail on a cluster that was already printing "must be
vacuumed" warnings.  Follow the warning's instructions to clear the
warning, then run pg_upgrade again.  One can still, peacefully consume
98% of XIDs or mxacts, so DBAs need not change routine VACUUM settings.

Discussion: https://postgr.es/m/20200621083513.GA3074645@rfd.leadboat.com

This allows AM-specific knowledge to be applied during creation of
pg_amop and pg_amproc entries.  Specifically, the AM knows better than
core code which entries to consider as required or optional.  Giving
the latter entries the appropriate sort of dependency allows them to
be dropped without taking out the whole opclass or opfamily; which
is something we'd like to have to correct obsolescent entries in
extensions.

This callback also opens the door to performing AM-specific validity
checks during opclass creation, rather than hoping than an opclass
developer will remember to test with "amvalidate".  For the most part
I've not actually added any such checks yet; that can happen in a
follow-on patch.  (Note that we shouldn't remove any tests from
"amvalidate", as those are still needed to cross-check manually
constructed entries in the initdb data.  So adding tests to
"amadjustmembers" will be somewhat duplicative, but it seems like
a good idea anyway.)

Patch by me, reviewed by Alexander Korotkov, Hamid Akhtar, and
Anastasia Lubennikova.

Discussion: https://postgr.es/m/4578.1565195302@sss.pgh.pa.us

This avoids lseek() system calls at every SLRU I/O, as was
done for relation files in commit c24dcd0c.
Reviewed-by: Ashwin Agrawal <aagrawal@pivotal.io>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKG%2Biqke4uTRFj8D8uEUUgj%2BRokPSp%2BCWM6YYzaaamG9Wvg%40mail.gmail.com
Discussion: https://postgr.es/m/CA%2BhUKGJ%2BoHhnvqjn3%3DHro7xu-YDR8FPr0FL6LF35kHRX%3D_bUzg%40mail.gmail.com

When doing multiple insertions in pg_shdepend for the copy of
dependencies from a template database in CREATE DATABASE, the same
number of slots would have been created and used all the time.  As the
number of items to insert is not known in advance, this makes most of
the slots created for nothing.  This improves the slot handling so as
slot creation only happens when needed, minimizing the overhead of the
operation.

Author: Michael Paquier
Reviewed-by: Daniel Gustafsson
Discussion: https://postgr.es/m/20200731024148.GB3317@paquier.xyz

When reading the code it's not obvious when one should prefer dynahash
over simplehash and vice-versa, so, for programmer-friendliness, add
comments to inform that decision.

Show sample simplehash method signatures.

Author: James Coleman <jtc331@gmail.com>
Discussion: https://postgr.es/m/CAAaqYe_dOF39gAJ8rL-a3YO3Qo96MHMRQ2whFjK5ZcU6YvMQSA%40mail.gmail.com

Commit eba77534 fixed an amcheck false positive bug involving
inconsistencies in TOAST input state between table and index.  A test
case was added that verified that such an inconsistency didn't result in
a spurious corruption related error.

Test coverage from the test was accidentally lost by commit 501e41dd,
which propagated ALTER TABLE ...  SET STORAGE attstorage state to
indexes.  This broke the test because the test specifically relied on
attstorage not being propagated.  This artificially forced there to be
index tuples whose datums were equivalent to the datums in the heap
without the datums actually being bitwise equal.

Fix this by updating pg_attribute directly instead.  Commit 501e41dd
made similar changes to a test_decoding TOAST-related test case which
made the same assumption, but overlooked the amcheck test case.

Backpatch: 11-, just like commit eba77534 (and commit 501e41dd).

If a base type supports typmods, its array type does too, with the
same interpretation.  Hence changes in pg_type.typmodin/typmodout
must be propagated to the array type.

While here, improve AlterTypeRecurse to not recurse to domains if
there is nothing we'd need to change.

Oversight in fe30e7eb.  Back-patch to v13 where that came in.

The new hlCover() algorithm that I introduced in commit c9b0c678
turns out to potentially take O(N^2) or worse time on long documents,
if there are many occurrences of individual query words but few or no
substrings that actually satisfy the query.  (One way to hit this
behavior is with a "common_word & rare_word" type of query.)  This
seems unavoidable given the original goal of checking every substring
of the document, so we have to back off that idea.  Fortunately, it
seems unlikely that anyone would really want headlines spanning all of
a long document, so we can avoid the worse-than-linear behavior by
imposing a maximum length of substring that we'll consider.

For now, just hard-wire that maximum length as a multiple of max_words
times max_fragments.  Perhaps at some point somebody will argue for
exposing it as a ts_headline parameter, but I'm hesitant to make such
a feature addition in a back-patched bug fix.

I also noted that the hlFirstIndex() function I'd added in that
commit was unnecessarily stupid: it really only needs to check whether
a HeadlineWordEntry's item pointer is null or not.  This wouldn't make
all that much difference in typical cases with queries having just
a few terms, but a cycle shaved is a cycle earned.

In addition, add a CHECK_FOR_INTERRUPTS call in TS_execute_recurse.
This ensures that hlCover's loop is cancellable if it manages to take
a long time, and it may protect some other TS_execute callers as well.

Back-patch to 9.6 as the previous commit was.  I also chose to add the
CHECK_FOR_INTERRUPTS call to 9.5.  The old hlCover() algorithm seems
to avoid the O(N^2) behavior, at least on the test case I tried, but
nonetheless it's not very quick on a long document.

Per report from Stephen Frost.

Discussion: https://postgr.es/m/20200724160535.GW12375@tamriel.snowman.net

Per build farm.

Discussion: https://postgr.es/m/20200731062626.GD3317%40paquier.xyz

Create an optional region in the main shared memory segment that can be
used to acquire and release "fast" DSM segments, and can benefit from
huge pages allocated at cluster startup time, if configured.  Fall back
to the existing mechanisms when that space is full.  The size is
controlled by a new GUC min_dynamic_shared_memory, defaulting to 0.

Main region DSM segments initially contain whatever garbage the memory
held last time they were used, rather than zeroes.  That change revealed
that DSA areas failed to initialize themselves correctly in memory that
wasn't zeroed first, so fix that problem.

Discussion: https://postgr.es/m/CA%2BhUKGLAE2QBv-WgGp%2BD9P_J-%3Dyne3zof9nfMaqq1h3EGHFXYQ%40mail.gmail.com

local_blks_dirtied tracks the number of local blocks dirtied, not shared
ones.

Author: Kirk Jamison
Discussion: https://postgr.es/m/OSBPR01MB2341760686DC056DE89D2AB9EF710@OSBPR01MB2341.jpnprd01.prod.outlook.com

Avoid repeatedly calling lseek(SEEK_END) during recovery by caching
the size of each fork.  For now, we can't use the same technique in
other processes, because we lack a shared invalidation mechanism.

Do this by generalizing the pre-existing caching used by FSM and VM
to support all forks.

Discussion: https://postgr.es/m/CAEepm%3D3SSw-Ty1DFcK%3D1rU-K6GSzYzfdD4d%2BZwapdN7dTa6%3DnQ%40mail.gmail.com

For pg_attribute, this allows to insert at once a full set of attributes
for a relation (roughly 15% of WAL reduction in extreme cases).  For
pg_shdepend, this reduces the work done when creating new shared
dependencies from a database template.  The number of slots used for the
insertion is capped at 64kB of data inserted for both, depending on the
number of items to insert and the length of the rows involved.

More can be done for other catalogs, like pg_depend.  This part requires
a different approach as the number of slots to use depends also on the
number of entries discarded as pinned dependencies.  This is also
related to the rework or dependency handling for ALTER TABLE and CREATE
TABLE, mainly.

Author: Daniel Gustafsson
Reviewed-by: Andres Freund, Michael Paquier
Discussion: https://postgr.es/m/20190213182737.mxn6hkdxwrzgxk35@alap3.anarazel.de

In "High Availability, Load Balancing, and Replication" chapter,
certain descriptions of Pgpool-II were not correct at this point.  It
does not need conflict resolution. Also "Multiple-Server Parallel
Query Execution" is not supported anymore.

Discussion: https://postgr.es/m/20200726.230128.53842489850344110.t-ishii%40sraoss.co.jp
Author: Tatsuo Ishii
Reviewed-by: Bruce Momjian
Backpatch-through: 9.5

Using pg_leftmost_one_post32() yields substantial performance benefits.

Backpatching to version 13 because HLL is used for HashAgg
improvements in 9878b643, which was also backpatched to 13.

Reviewed-by: Peter Geoghegan
Discussion: https://postgr.es/m/CAH2-WzkGvDKVDo+0YvfvZ+1CE=iCi88DCOGFF3i1hTGGaxcKPw@mail.gmail.com
Backpatch-through: 13