Some conditions resulted in "return" directly out of a PG_TRY block,
which left the exception stack dangling, and to add insult to injury
failed to restore the state of set_latch_on_sigusr1.
This is a bug only in 9.5; in HEAD it was accidentally fixed by commit
db0f6cad4, which removed the surrounding PG_TRY block. However, I (tgl)
chose to apply the patch to HEAD as well, because the old coding was
gratuitously different from WaitForBackgroundWorkerStartup(), and there
would indeed have been no bug if it were done like that to start with.
Dmitry Ivanov
Discussion: <1637882.WfYN5gPf1A@abook>
We don't want postmaster child processes to contain a copy of the
postmaster's PostmasterContext. That would be a waste of memory at least,
and at worst a security issue, since there are copies of the semi-sensitive
pg_hba and pg_ident data in there. All other child process types delete
the PostmasterContext after forking, but the original coding of the
background worker patch (commit da07a1e85) did not do so. It appears
that the only reason for that was to avoid copying the bgworker's
MyBgworkerEntry out of that context; but the couple of additional
statements needed to do so are hardly good justification for it. Hence,
copy that data and then clear the context as other child processes do.
Because this patch changes the memory context in which a bgworker function
gains control, back-patching it would be a bit risky, so we won't fix this
in back branches. The "security" complaint is pretty thin anyway for
generic bgworkers; only with the introduction of parallel query is there
any question of running untrusted code in a bgworker process.
Discussion: <14111.1470082717@sss.pgh.pa.us>
This is apparently harmless on Windows, but on Unix it results in an
assertion failure. We'd not noticed because this code doesn't get
used on Unix unless you build with -DEXEC_BACKEND. Bug was evidently
introduced by sloppy refactoring in commit 31c453165.
Thomas Munro
Discussion: <CAEepm=1VOnbVx4wsgQFvj94hu9jVt2nVabCr7QiooUSvPJXkgQ@mail.gmail.com>
Commit 0b0baf262 et al made this case print "(null)" on the grounds that
that's what happened on platforms that didn't crash. But neither behavior
was actually intentional. What we should print is just an empty string,
for compatibility with the behavior of SHOW and other ways of examining
string GUCs. Those code paths don't distinguish NULL from empty strings,
so we should not here either. Per gripe from Alain Radix.
Like the previous patch, back-patch to 9.2 where -C option was introduced.
Discussion: <CA+YdpwxPUADrmxSD7+Td=uOshMB1KkDN7G7cf+FGmNjjxMhjbw@mail.gmail.com>
Emit "(null)" instead, which was the behavior all along on platforms
that don't crash, eg OS X. Per report from Jehan-Guillaume de Rorthais.
Back-patch to 9.2 where -C option was introduced.
Michael Paquier
Report: <20160615204036.2d35d86a@firost>
We disable statement_timeout and lock_timeout during dump and restore, to
prevent any global settings that might exist from breaking routine backups.
Commit c6dda1f48 should have added idle_in_transaction_session_timeout to
that list, but failed to.
Another place where these timeouts get turned off is autovacuum. While
I doubt an idle timeout could fire there, it seems better to be safe than
sorry.
pg_dump issue noted by Bernd Helmle, the other one found by grepping.
Report: <352F9B77DB5D3082578D17BB@eje.land.credativ.lan>
If we ANALYZE only selected columns of a table, we should not postpone
auto-analyze because of that; other columns may well still need stats
updates. As committed, the counter is left alone if a column list is
given, whether or not it includes all analyzable columns of the table.
Per complaint from Tomasz Ostrowski.
It's been like this a long time, so back-patch to all supported branches.
Report: <ef99c1bd-ff60-5f32-2733-c7b504eb960c@ato.waw.pl>
The original intent in the stats collector was that we should not write out
stats data oftener than every PGSTAT_STAT_INTERVAL msec. Backends will not
make requests at all if they see the existing data is newer than that, and
the stats collector is supposed to disregard requests having a cutoff_time
older than its most recently written data, so that close-together requests
don't result in multiple writes. But the latter part of that got broken
in commit 187492b6c2, so that if two backends concurrently decide
the existing stats are too old, the collector would write the data twice.
(In principle the collector's logic would still merge requests as long as
the second one arrives before we've actually written data ... but since
the message collection loop would write data immediately after processing
a single inquiry message, that never happened in practice, and in any case
the window in which it might work would be much shorter than
PGSTAT_STAT_INTERVAL.)
To fix, improve pgstat_recv_inquiry so that it checks whether the cutoff
time is too old, and doesn't add a request to the queue if so. This means
that we do not need DBWriteRequest.request_time, because the decision is
taken before making a queue entry. And that means that we don't really
need the DBWriteRequest data structure at all; an OID list of database
OIDs will serve and allow removal of some rather verbose and crufty code.
In passing, improve the comments in this area, which have been rather
neglected. Also change backend_read_statsfile so that it's not silently
relying on MyDatabaseId to have some particular value in the autovacuum
launcher process. It accidentally worked as desired because MyDatabaseId
is zero in that process; but that does not seem like a dependency we want,
especially with no documentation about it.
Although this patch is mine, it turns out I'd rediscovered a known bug,
for which Tomas Vondra had already submitted a patch that's functionally
equivalent to the non-cosmetic aspects of this patch. Thanks to Tomas
for reviewing this version.
Back-patch to 9.3 where the bug was introduced.
Prior-Discussion: <1718942738eb65c8407fcd864883f4c8@fuzzy.cz>
Patch: <4625.1464202586@sss.pgh.pa.us>
Ever since we split the statistics collector's reports into per-database
files (commit 187492b6c2), backends have been seeing stale statistics
for shared catalogs. This is because the inquiry message only prompts the
collector to write the per-database file for the requesting backend's own
database. Stats for shared catalogs are in a separate file for "DB 0",
which didn't get updated.
In normal operation this was partially masked by the fact that the
autovacuum launcher would send an inquiry message at least once per
autovacuum_naptime that asked for "DB 0"; so the shared-catalog stats would
never be more than a minute out of date. However the problem becomes very
obvious with autovacuum disabled, as reported by Peter Eisentraut.
To fix, redefine the semantics of inquiry messages so that both the
specified DB and DB 0 will be dumped. (This might seem a bit inefficient,
but we have no good way to know whether a backend's transaction will look
at shared-catalog stats, so we have to read both groups of stats whenever
we request stats. Sending two inquiry messages would definitely not be
better.)
Back-patch to 9.3 where the bug was introduced.
Report: <56AD41AC.1030509@gmx.net>
The table-skipping logic in autovacuum would fail to consider that
multiple workers could be processing the same shared catalog in
different databases. This normally wouldn't be a problem: firstly
because autovacuum workers not for wraparound would simply ignore tables
in which they cannot acquire lock, and secondly because most of the time
these tables are small enough that even if multiple for-wraparound
workers are stuck in the same catalog, they would be over pretty
quickly. But in cases where the catalogs are severely bloated it could
become a problem.
Backpatch all the way back, because the problem has been there since the
beginning.
Reported by Ondřej Světlík
Discussion: https://www.postgresql.org/message-id/572B63B1.3030603%40flexibee.euhttps://www.postgresql.org/message-id/572A1072.5080308%40flexibee.eu
Now that pg_dump will properly dump out any ACL changes made to
functions which exist in pg_catalog, switch to using the GRANT system
to manage access to those functions.
This means removing 'if (!superuser()) ereport()' checks from the
functions themselves and then REVOKEing EXECUTE right from 'public' for
these functions in system_views.sql.
Reviews by Alexander Korotkov, Jose Luis Tallon
If archive_timeout > 0 we should avoid logging XLOG_RUNNING_XACTS if idle.
Bug 13685 reported by Laurence Rowe, investigated in detail by Michael Paquier,
though this is not his proposed fix.
20151016203031.3019.72930@wrigleys.postgresql.org
Simple non-invasive patch to allow later backpatch to 9.4 and 9.5
The distinction between "archive" and "hot_standby" existed only because
at the time "hot_standby" was added, there was some uncertainty about
stability. This is now a long time ago. We would like to move forward
with simplifying the replication configuration, but this distinction is
in the way, because a primary server cannot tell (without asking a
standby or predicting the future) which one of these would be the
appropriate level.
Pick a new name for the combined setting to make it clearer that it
covers all (non-logical) backup and replication uses. The old values
are still accepted but are converted internally.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Reviewed-by: David Steele <david@pgmasters.net>
Vinayak Pokale provided a patch for a copy-and-paste error in a
comment. I noticed that I'd use the word "automatically" nearby where
I meant to talk about things being "atomic". Rahila Syed spotted a
misplaced counter update. Fix all that stuff.
There's a lot more that could be done here yet - in particular, this
reports only very coarse-grained information about the index vacuuming
phase - but even as it stands, the new pg_stat_progress_vacuum can
tell you quite a bit about what a long-running vacuum is actually
doing.
Amit Langote and Robert Haas, based on earlier work by Vinayak Pokale
and Rahila Syed.
Currently writes to the main data files of postgres all go through the
OS page cache. This means that some operating systems can end up
collecting a large number of dirty buffers in their respective page
caches. When these dirty buffers are flushed to storage rapidly, be it
because of fsync(), timeouts, or dirty ratios, latency for other reads
and writes can increase massively. This is the primary reason for
regular massive stalls observed in real world scenarios and artificial
benchmarks; on rotating disks stalls on the order of hundreds of seconds
have been observed.
On linux it is possible to control this by reducing the global dirty
limits significantly, reducing the above problem. But global
configuration is rather problematic because it'll affect other
applications; also PostgreSQL itself doesn't always generally want this
behavior, e.g. for temporary files it's undesirable.
Several operating systems allow some control over the kernel page
cache. Linux has sync_file_range(2), several posix systems have msync(2)
and posix_fadvise(2). sync_file_range(2) is preferable because it
requires no special setup, whereas msync() requires the to-be-flushed
range to be mmap'ed. For the purpose of flushing dirty data
posix_fadvise(2) is the worst alternative, as flushing dirty data is
just a side-effect of POSIX_FADV_DONTNEED, which also removes the pages
from the page cache. Thus the feature is enabled by default only on
linux, but can be enabled on all systems that have any of the above
APIs.
While desirable and likely possible this patch does not contain an
implementation for windows.
With the infrastructure added, writes made via checkpointer, bgwriter
and normal user backends can be flushed after a configurable number of
writes. Each of these sources of writes controlled by a separate GUC,
checkpointer_flush_after, bgwriter_flush_after and backend_flush_after
respectively; they're separate because the number of flushes that are
good are separate, and because the performance considerations of
controlled flushing for each of these are different.
A later patch will add checkpoint sorting - after that flushes from the
ckeckpoint will almost always be desirable. Bgwriter flushes are most of
the time going to be random, which are slow on lots of storage hardware.
Flushing in backends works well if the storage and bgwriter can keep up,
but if not it can have negative consequences. This patch is likely to
have negative performance consequences without checkpoint sorting, but
unfortunately so has sorting without flush control.
Discussion: alpine.DEB.2.10.1506011320000.28433@sto
Author: Fabien Coelho and Andres Freund
Earlier version committed in 9.0 caused spurious waits in some cases.
New infrastructure for lock waits in 9.3 used to correct and improve this.
Jeff Janes based upon a proposal by Simon Riggs, who also reviewed
Additional review comments from Amit Kapila
When a process is waiting for a heavyweight lock, we will now indicate
the type of heavyweight lock for which it is waiting. Also, you can
now see when a process is waiting for a lightweight lock - in which
case we will indicate the individual lock name or the tranche, as
appropriate - or for a buffer pin.
Amit Kapila, Ildus Kurbangaliev, reviewed by me. Lots of helpful
discussion and suggestions by many others, including Alexander
Korotkov, Vladimir Borodin, and many others.
Renaming a file using rename(2) is not guaranteed to be durable in face
of crashes. Use the previously added durable_rename()/durable_link_or_rename()
in various places where we previously just renamed files.
Most of the changed call sites are arguably not critical, but it seems
better to err on the side of too much durability. The most prominent
known case where the previously missing fsyncs could cause data loss is
crashes at the end of a checkpoint. After the actual checkpoint has been
performed, old WAL files are recycled. When they're filled, their
contents are fdatasynced, but we did not fsync the containing
directory. An OS/hardware crash in an unfortunate moment could then end
up leaving that file with its old name, but new content; WAL replay
would thus not replay it.
Reported-By: Tomas Vondra
Author: Michael Paquier, Tomas Vondra, Andres Freund
Discussion: 56583BDD.9060302@2ndquadrant.com
Backpatch: All supported branches
Using this facility, any utility command can report the target relation
upon which it is operating, if there is one, and up to 10 64-bit
counters; the intent of this is that users should be able to figure out
what a utility command is doing without having to resort to ugly hacks
like attaching strace to a backend.
As a demonstration, this adds very crude reporting to lazy vacuum; we
just report the target relation and nothing else. A forthcoming patch
will make VACUUM report a bunch of additional data that will make this
much more interesting. But this gets the basic framework in place.
Vinayak Pokale, Rahila Syed, Amit Langote, Robert Haas, reviewed by
Kyotaro Horiguchi, Jim Nasby, Thom Brown, Masahiko Sawada, Fujii Masao,
and Masanori Oyama.
Commit 4de82f7d7 increased the WAL flush rate, mainly to increase the
likelihood that hint bits can be set quickly. More quickly set hint bits
can reduce contention around the clog et al. But unfortunately the
increased flush rate can have a significant negative performance impact,
I have measured up to a factor of ~4. The reason for this slowdown is
that if there are independent writes to the underlying devices, for
example because shared buffers is a lot smaller than the hot data set,
or because a checkpoint is ongoing, the fdatasync() calls force cache
flushes to be emitted to the storage.
This is achieved by flushing WAL only if the last flush was longer than
wal_writer_delay ago, or if more than wal_writer_flush_after (new GUC)
unflushed blocks are pending. Based on some tests the default for
wal_writer_delay is 1MB, which seems to work well both on SSD and
rotational media.
To avoid negative performance impact due to 4de82f7d7 an earlier
commit (db76b1e) made SetHintBits() more likely to succeed; preventing
performance regressions in the pgbench tests I performed.
Discussion: 20160118163908.GW10941@awork2.anarazel.de
Further investigation says that there may be some slow operations after
we've finished ShutdownXLOG(), so add some more log messages to try to
isolate that. This is all temporary code too.
This is a quick hack, due to be reverted when its purpose has been served,
to try to gather information about why some of the buildfarm critters
regularly fail with "postmaster does not shut down" complaints. Maybe they
are just really overloaded, but maybe something else is going on. Hence,
instrument pg_ctl to print the current time when it starts waiting for
postmaster shutdown and when it gives up, and add a lot of logging of the
current time in the server's checkpoint and shutdown code paths.
No attempt has been made to make this pretty. I'm not even totally sure
if it will build on Windows, but we'll soon find out.
The previous RequestAddinLWLocks() method had several disadvantages.
First, the locks would be in the main tranche; we've recently decided
that it's useful for LWLocks used for separate purposes to have
separate tranche IDs. Second, there wasn't any correlation between
what code called RequestAddinLWLocks() and what code called
LWLockAssign(); when multiple modules are in use, it could become
quite difficult to troubleshoot problems where LWLockAssign() ran out
of locks. To fix, create a concept of named LWLock tranches which
can be used either by extension or by core code.
Amit Kapila and Robert Haas
Insert sd_notify() calls at server start and stop for integration with
systemd. This allows the use of systemd service units of type "notify",
which greatly simplifies the systemd configuration.
Reviewed-by: Pavel Stěhule <pavel.stehule@gmail.com>
We entirely randomly chose to initialize port->remote_host just after
printing the log_connections message, when we could perfectly well do it
just before, allowing %h and %r to work for that message. Per gripe from
Artem Tomyuk.
This patch reduces pg_am to just two columns, a name and a handler
function. All the data formerly obtained from pg_am is now provided
in a C struct returned by the handler function. This is similar to
the designs we've adopted for FDWs and tablesample methods. There
are multiple advantages. For one, the index AM's support functions
are now simple C functions, making them faster to call and much less
error-prone, since the C compiler can now check function signatures.
For another, this will make it far more practical to define index access
methods in installable extensions.
A disadvantage is that SQL-level code can no longer see attributes
of index AMs; in particular, some of the crosschecks in the opr_sanity
regression test are no longer possible from SQL. We've addressed that
by adding a facility for the index AM to perform such checks instead.
(Much more could be done in that line, but for now we're content if the
amvalidate functions more or less replace what opr_sanity used to do.)
We might also want to expose some sort of reporting functionality, but
this patch doesn't do that.
Alexander Korotkov, reviewed by Petr Jelínek, and rather heavily
editorialized on by me.
Up until now, the total amount of data that could be passed to a
background worker at startup was one datum, which can be a small as
4 bytes on some systems. That's enough to pass a dsm_handle or an
array index, but not much else. Add a bgw_extra flag to the
BackgroundWorker struct, allowing up to 128 bytes to be passed to
a new worker on any platform.
Use this to fix a problem I recently discovered with the parallel
context machinery added in 9.5: the master assigns each worker an
array index, and each worker subsequently assigns itself an array
index, and there's nothing to guarantee that the two sets of indexes
match, leading to chaos.
Normally, I would not back-patch the change to add bgw_extra, since it
is basically a feature addition. However, since 9.5 is still in beta
and there seems to be no other sensible way to repair the broken
parallel context machinery, back-patch to 9.5. Existing background
worker code can ignore the bgw_extra field without a problem, but
might need to be recompiled since the structure size has changed.
Report and patch by me. Review by Amit Kapila.
Postmaster child processes that aren't supposed to be attached to shared
memory were not bothering to close the shared memory mapping handle they
inherit from the postmaster process. That's mostly harmless, since the
handle vanishes anyway when the child process exits -- but the syslogger
process, if used, doesn't get killed and restarted during recovery from a
backend crash. That meant that Windows doesn't see the shared memory
mapping as becoming free, so it doesn't delete it and the postmaster is
unable to create a new one, resulting in failure to recover from crashes
whenever logging_collector is turned on.
Per report from Dmitry Vasilyev. It's a bit astonishing that we'd not
figured this out long ago, since it's been broken from the very beginnings
of out native Windows support; probably some previously-unexplained trouble
reports trace to this.
A secondary problem is that on Cygwin (perhaps only in older versions?),
exec() may not detach from the shared memory segment after all, in which
case these child processes did remain attached to shared memory, posing
the risk of an unexpected shared memory clobber if they went off the rails
somehow. That may be a long-gone bug, but we can deal with it now if it's
still live, by detaching within the infrastructure introduced here to deal
with closing the handle.
Back-patch to all supported branches.
Tom Lane and Amit Kapila
This flag has proven to be a recipe for bugs, and it doesn't seem like
it can really buy anything in terms of performance. So let's just
*always* set the process latch when we receive SIGUSR1 instead of
trying to do it only when needed.
Per my recent proposal on pgsql-hackers.
For some reason, neither of the compilers I usually use noticed the
uninitialized-variable problem I introduced in commit 7e2a18a916.
That's hardly a good enough excuse though. Committing with brown paper bag
on head.
In addition to putting the operations in the right order, move the
declaration of "now" inside the loop; there's no need for it to be
outside, and that does wake up older gcc enough to notice any similar
future problem.
Back-patch to 9.4; earlier versions lack the time-to-SIGKILL stanza
so there's no bug.
The postmaster now checks every minute or so (worst case, at most two
minutes) that postmaster.pid is still there and still contains its own PID.
If not, it performs an immediate shutdown, as though it had received
SIGQUIT.
The original goal behind this change was to ensure that failed buildfarm
runs would get fully cleaned up, even if the test scripts had left a
postmaster running, which is not an infrequent occurrence. When the
buildfarm script removes a test postmaster's $PGDATA directory, its next
check on postmaster.pid will fail and cause it to exit. Previously, manual
intervention was often needed to get rid of such orphaned postmasters,
since they'd block new test postmasters from obtaining the expected socket
address.
However, by checking postmaster.pid and not something else, we can provide
additional robustness: manual removal of postmaster.pid is a frequent DBA
mistake, and now we can at least limit the damage that will ensue if a new
postmaster is started while the old one is still alive.
Back-patch to all supported branches, since we won't get the desired
improvement in buildfarm reliability otherwise.
Prior to commit 0709b7ee72, access to
variables within a spinlock-protected critical section had to be done
through a volatile pointer, but that should no longer be necessary.
This continues work begun in df4077cda2
and 6ba4ecbf47.
Thomas Munro and Michael Paquier
This commit makes postmaster forcibly remove the files signaling
a standby promotion request. Otherwise, the existence of those files
can trigger a promotion too early, whether a user wants that or not.
This removal of files is usually unnecessary because they can exist
only during a few moments during a standby promotion. However
there is a race condition: if pg_ctl promote is executed and creates
the files during a promotion, the files can stay around even after
the server is brought up to new master. Then, if new standby starts
by using the backup taken from that master, the files can exist
at the server startup and should be removed in order to avoid
an unexpected promotion.
Back-patch to 9.1 where promote signal file was introduced.
Problem reported by Feike Steenbergen.
Original patch by Michael Paquier, modified by me.
Discussion: 20150528100705.4686.91426@wrigleys.postgresql.org
Previously, if one background worker registered another background
worker and set bgw_notify_pid while for the second background worker,
it would not receive notifications from the postmaster unless, at the
time the "parent" was registered, BGWORKER_BACKEND_DATABASE_CONNECTION
was set.
To fix, instead instead of including only those background workers that
requested database connections in the postmater's BackendList, include
them all. There doesn't seem to be any reason not do this, and indeed
it removes a significant amount of duplicated code. The other option
is to make PostmasterMarkPIDForWorkerNotify look at BackgroundWorkerList
in addition to BackendList, but that adds more code duplication instead
of getting rid of it.
Patch by me. Review and testing by Ashutosh Bapat.
Commit c9b0cbe98b accidentally broke the
order of operations during postmaster shutdown: it resulted in removing
the per-socket lockfiles after, not before, postmaster.pid. This creates
a race-condition hazard for a new postmaster that's started immediately
after observing that postmaster.pid has disappeared; if it sees the
socket lockfile still present, it will quite properly refuse to start.
This error appears to be the explanation for at least some of the
intermittent buildfarm failures we've seen in the pg_upgrade test.
Another problem, which has been there all along, is that the postmaster
has never bothered to close() its listen sockets, but has just allowed them
to close at process death. This creates a different race condition for an
incoming postmaster: it might be unable to bind to the desired listen
address because the old postmaster is still incumbent. This might explain
some odd failures we've seen in the past, too. (Note: this is not related
to the fact that individual backends don't close their client communication
sockets. That behavior is intentional and is not changed by this patch.)
Fix by adding an on_proc_exit function that closes the postmaster's ports
explicitly, and (in 9.3 and up) reshuffling the responsibility for where
to unlink the Unix socket files. Lock file unlinking can stay where it
is, but teach it to unlink the lock files in reverse order of creation.
Fix additional bogosity in commit 9029f4b374. Include the
BackendSslStatusBuffer in the BackendStatusShmemSize calculation,
avoid ugly and error-prone casts to char* and back, put related
code stanzas into a consistent order (and fix a couple of previous
instances of that sin). All cosmetic except for the size oversight.
Ordinarily, a failure (unexpected exit status) of the startup subprocess
should be considered fatal, so the postmaster should just close up shop
and quit. However, if we sent the startup process a SIGQUIT or SIGKILL
signal, the failure is hardly "unexpected", and we should attempt restart;
this is necessary for recovery from ordinary backend crashes in hot-standby
scenarios. I attempted to implement the latter rule with a two-line patch
in commit 442231d7f7, but it now emerges that
that patch was a few bricks shy of a load: it failed to distinguish the
case of a signaled startup process from the case where the new startup
process crashes before reaching database consistency. That resulted in
infinitely respawning a new startup process only to have it crash again.
To handle this properly, we really must track whether we have sent the
*current* startup process a kill signal. Rather than add yet another
ad-hoc boolean to the postmaster's state, I chose to unify this with the
existing RecoveryError flag into an enum tracking the startup process's
state. That seems more consistent with the postmaster's general state
machine design.
Back-patch to 9.0, like the previous patch.
When archive recovery and restartpoints were initially introduced,
checkpoint_segments was ignored on the grounds that the files restored from
archive don't consume any space in the recovery server. That was changed in
later releases, but even then it was arguably a feature rather than a bug,
as performing restartpoints as often as checkpoints during normal operation
might be excessive, but you might nevertheless not want to waste a lot of
space for pre-allocated WAL by setting checkpoint_segments to a high value.
But now that we have separate min_wal_size and max_wal_size settings, you
can bound WAL usage with max_wal_size, and still avoid consuming excessive
space usage by setting min_wal_size to a lower value, so that argument is
moot.
There are still some issues with actually limiting the space usage to
max_wal_size: restartpoints in recovery can only start after seeing the
checkpoint record, while a checkpoint starts flushing buffers as soon as
the redo-pointer is set. Restartpoint is paced to happen at the same
leisurily speed, determined by checkpoint_completion_target, as checkpoints,
but because they are started later, max_wal_size can be exceeded by upto
one checkpoint cycle's worth of WAL, depending on
checkpoint_completion_target. But that seems better than not trying at all,
and max_wal_size is a soft limit anyway.
The documentation already claimed that max_wal_size is obeyed in recovery,
so this just fixes the behaviour to match the docs. However, add some
weasel-words there to mention that max_wal_size may well be exceeded by
some amount in recovery.
Tom Lane