After a timeline switch, we would leave behind recycled WAL segments that
are in the future, but on the old timeline. After promotion, and after they
become old enough to be recycled again, we would notice that they don't have
a .ready or .done file, create a .ready file for them, and archive them.
That's bogus, because the files contain garbage, recycled from an older
timeline (or prealloced as zeros). We shouldn't archive such files.
This could happen when we're following a timeline switch during replay, or
when we switch to new timeline at end-of-recovery.
To fix, whenever we switch to a new timeline, scan the data directory for
WAL segments on the old timeline, but with a higher segment number, and
remove them. Those don't belong to our timeline history, and are most
likely bogus recycled or preallocated files. They could also be valid files
that we streamed from the primary ahead of time, but in any case, they're
not needed to recover to the new timeline.
Similarly to previous fix 9b8d478, commit 2c03216 has switched
XLogReaderAllocate() to use a set of palloc calls instead of malloc,
causing any callers of this function to fail with an error instead of
receiving a NULL pointer in case of out-of-memory error. Fix this by
using palloc_extended with MCXT_ALLOC_NO_OOM that will safely return
NULL in case of an OOM.
Michael Paquier, slightly modified by me.
In 83ff1618 we defined integer limits iff they're not provided by the
system. That turns out not to be the greatest idea because there's
different ways some datatypes can be represented. E.g. on OSX PG's 64bit
datatype will be a 'long int', but OSX unconditionally uses 'long
long'. That disparity then can lead to warnings, e.g. around printf
formats.
One way to fix that would be to back int64 using stdint.h's
int64_t. While a good idea it's not that easy to implement. We would
e.g. need to include stdint.h in our external headers, which we don't
today. Also computing the correct int64 printf formats in that case is
nontrivial.
Instead simply prefix the integer limits with PG_ and define them
unconditionally. I've adjusted all the references to them in code, but
not the ones in comments; the latter seems unnecessary to me.
Discussion: 20150331141423.GK4878@alap3.anarazel.de
Several submitted and even committed patches have run into the problem
that C89, our baseline, does not provide minimum/maximum values for
various integer datatypes. C99's stdint.h does, but we can't rely on
it.
Several parts of the code defined limits locally, so instead centralize
the definitions to c.h.
This patch also changes the more obvious usages of literal limit values;
there's more places that could be changed, but it's less clear whether
it's beneficial to change those.
Author: Andrew Gierth
Discussion: 87619tc5wc.fsf@news-spur.riddles.org.uk
Recovery delays are implemented by waiting on a latch, and latches take
milliseconds as a parameter. The required amount of waiting was computed
using microsecond resolution though and the wait loop's abort condition
was checking the delay in microseconds as well. This could lead to
short spurts of busy looping when the overall wait time was below a
millisecond, but above 0 microseconds.
Instead just formulate the wait loop's abort condition in millisecond
granularity as well. Given that that's recovery_min_apply_delay
resolution, it seems harmless to not wait for less than a millisecond.
Backpatch to 9.4 where recovery_min_apply_delay was introduced.
Discussion: 20150323141819.GH26995@alap3.anarazel.de
Since 465883b0a two versions of commit records have existed. A compact
version that was used when no cache invalidations, smgr unlinks and
similar were needed, and a full version that could deal with all
that. Additionally the full version was embedded into twophase commit
records.
That resulted in a measurable reduction in the size of the logged WAL in
some workloads. But more recently additions like logical decoding, which
e.g. needs information about the database something was executed on,
made it applicable in fewer situations. The static split generally made
it hard to expand the commit record, because concerns over the size made
it hard to add anything to the compact version.
Additionally it's not particularly pretty to have twophase.c insert
RM_XACT records.
Rejigger things so that the commit and abort records only have one form
each, including the twophase equivalents. The presence of the various
optional (in the sense of not being in every record) pieces is indicated
by a bits in the 'xinfo' flag. That flag previously was not included in
compact commit records. To prevent an increase in size due to its
presence, it's only included if necessary; signalled by a bit in the
xl_info bits available for xact.c, similar to heapam.c's
XLOG_HEAP_OPMASK/XLOG_HEAP_INIT_PAGE.
Twophase commit/aborts are now the same as their normal
counterparts. The original transaction's xid is included in an optional
data field.
This means that commit records generally are smaller, except in the case
of a transaction with subtransactions, but no other special cases; the
increase there is four bytes, which seems acceptable given that the more
common case of not having subtransactions shrank. The savings are
especially measurable for twophase commits, which previously always used
the full version; but will in practice only infrequently have required
that.
The motivation for this work are not the space savings and and
deduplication though; it's that it makes it easier to extend commit
records with additional information. That's just a few lines of code
now; without impacting the common case where that information is not
needed.
Discussion: 20150220152150.GD4149@awork2.anarazel.de,
235610.92468.qm%40web29004.mail.ird.yahoo.com
Reviewed-By: Heikki Linnakangas, Simon Riggs
The introduction of min_wal_size & max_wal_size in 88e9823026 makes it
feasible to increase the default upper bound in checkpoint
size. Previously raising the default would lead to a increased disk
footprint, even if more segments weren't beneficial. The low default of
checkpoint size is one of common performance problem users have thus
increasing the default makes sense. Setups where the increase in
maximum disk usage is a problem will very likely have to run with a
modified configuration anyway.
Discussion: 54F4EFB8.40202@agliodbs.com,
CA+TgmoZEAgX5oMGJOHVj8L7XOkAe05Gnf45rP40m-K3FhZRVKg@mail.gmail.com
Author: Josh Berkus, after a discussion involving lots of people.
The new recovery_target_action (introduced in aedccb1f6/b8e33a85d4)
replaces it's functionality. Having both seems likely to cause more
confusion than it saves worry due to the incompatibility.
Discussion: 5484FC53.2060903@2ndquadrant.com
Author: Petr Jelinek
When newly-added GUC parameter, wal_compression, is on, the PostgreSQL server
compresses a full page image written to WAL when full_page_writes is on or
during a base backup. A compressed page image will be decompressed during WAL
replay. Turning this parameter on can reduce the WAL volume without increasing
the risk of unrecoverable data corruption, but at the cost of some extra CPU
spent on the compression during WAL logging and on the decompression during
WAL replay.
This commit changes the WAL format (so bumping WAL version number) so that
the one-byte flag indicating whether a full page image is compressed or not is
included in its header information. This means that the commit increases the
WAL volume one-byte per a full page image even if WAL compression is not used
at all. We can save that one-byte by borrowing one-bit from the existing field
like hole_offset in the header and using it as the flag, for example. But which
would reduce the code readability and the extensibility of the feature.
Per discussion, it's not worth paying those prices to save only one-byte, so we
decided to add the one-byte flag to the header.
This commit doesn't introduce any new compression algorithm like lz4.
Currently a full page image is compressed using the existing PGLZ algorithm.
Per discussion, we decided to use it at least in the first version of the
feature because there were no performance reports showing that its compression
ratio is unacceptably lower than that of other algorithm. Of course,
in the future, it's worth considering the support of other compression
algorithm for the better compression.
Rahila Syed and Michael Paquier, reviewed in various versions by myself,
Andres Freund, Robert Haas, Abhijit Menon-Sen and many others.
We allow this module to be turned off on restarts, so a restart time
check is enough to activate or deactivate the module; however, if there
is a standby replaying WAL emitted from a master which is restarted, but
the standby isn't, the state in the standby becomes inconsistent and can
easily be crashed.
Fix by activating and deactivating the module during WAL replay on
parameter change as well as on system start.
Problem reported by Fujii Masao in
http://www.postgresql.org/message-id/CAHGQGwFhJ3CnHo1CELEfay18yg_RA-XZT-7D8NuWUoYSZ90r4Q@mail.gmail.com
Author: Petr Jelínek
Instead of having a single knob (checkpoint_segments) that both triggers
checkpoints, and determines how many checkpoints to recycle, they are now
separate concerns. There is still an internal variable called
CheckpointSegments, which triggers checkpoints. But it no longer determines
how many segments to recycle at a checkpoint. That is now auto-tuned by
keeping a moving average of the distance between checkpoints (in bytes),
and trying to keep that many segments in reserve. The advantage of this is
that you can set max_wal_size very high, but the system won't actually
consume that much space if there isn't any need for it. The min_wal_size
sets a floor for that; you can effectively disable the auto-tuning behavior
by setting min_wal_size equal to max_wal_size.
The max_wal_size setting is now the actual target size of WAL at which a
new checkpoint is triggered, instead of the distance between checkpoints.
Previously, you could calculate the actual WAL usage with the formula
"(2 + checkpoint_completion_target) * checkpoint_segments + 1". With this
patch, you set the desired WAL usage with max_wal_size, and the system
calculates the appropriate CheckpointSegments with the reverse of that
formula. That's a lot more intuitive for administrators to set.
Reviewed by Amit Kapila and Venkata Balaji N.
Previously when the standby server failed to retrieve WAL files from any sources
(i.e., streaming replication, local pg_xlog directory or WAL archive), it always
waited for five seconds (hard-coded) before the next attempt. For example,
this is problematic in warm-standby because restore_command can fail
every five seconds even while new WAL file is expected to be unavailable for
a long time and flood the log files with its error messages.
This commit adds new parameter, wal_retrieve_retry_interval, to control that
wait time.
Alexey Vasiliev and Michael Paquier, reviewed by Andres Freund and me.
The flag is supposed to be copied from the record. Same issue with
track_commit_timestamps, but that's master-only.
Report and fix by Petr Jalinek. Backpatch to 9.4, where wal_log_hints was
added.
Since commit ba94518a, we used XLogFileOpen to open the next segment for
writing, but if the end-of-recovery happens exactly at a segment boundary,
the new segment might not exist yet. (Before ba94518a, XLogFileOpen was
correct, because we would open the previous segment if the switch happened
at the boundary.)
Instead of trying to create it if necessary, it's simpler to not bother
opening the segment at all. XLogWrite() will open or create it soon anyway,
after writing the checkpoint or end-of-recovery record.
Reported by Andres Freund.
At one point in the development of this feature, it was claimed that
allowing negative values would be useful to compensate for timezone
differences between master and slave servers. That was based on a mistaken
assumption that commit timestamps are recorded in local time; but of course
they're in UTC. Nor is a negative apply delay likely to be a sane way of
coping with server clock skew. However, the committed patch still treated
negative delays as doing something, and the timezone misapprehension
survived in the user documentation as well.
If recovery_min_apply_delay were a proper GUC we'd just set the minimum
allowed value to be zero; but for the moment it seems better to treat
negative settings as if they were zero.
In passing do some extra wordsmithing on the parameter's documentation,
including correcting a second misstatement that the parameter affects
processing of Restore Point records.
Issue noted by Michael Paquier, who also provided the code patch; doc
changes by me. Back-patch to 9.4 where the feature was introduced.
XLogFileInit() returns a file descriptor, which needs to be closed. The leak
was short-lived, since the startup process exits shortly afterwards, but it
was clearly a bug, nevertheless.
Per Coverity report.
We used time(null) to set a TimestampTz field, which gave bogus results.
Noticed while looking at pg_xlogdump output.
Backpatch to 9.3 and above, where the fast promotion was introduced.
Two changes:
1. When copying a WAL segment from old timeline to create the first segment
on the new timeline, only copy up to the point where the timeline switch
happens, and zero-fill the rest. This avoids corner cases where we might
think that the copied WAL from the previous timeline belong to the new
timeline.
2. If the timeline switch happens at a segment boundary, don't copy the
whole old segment to the new timeline. It's pointless, because it's 100%
identical to the old segment.
When starting up from a basebackup taken off a standby extra logic has
to be applied to compute the point where the data directory is
consistent. Normal base backups use a WAL record for that purpose, but
that isn't possible on a standby.
That logic had a error check ensuring that the cluster's control file
indicates being in recovery. Unfortunately that check was too strict,
disregarding the fact that the control file could also indicate that
the cluster was shut down while in recovery.
That's possible when the a cluster starting from a basebackup is shut
down before the backup label has been removed. When everything goes
well that's a short window, but when either restore_command or
primary_conninfo isn't configured correctly the window can get much
wider. That's because inbetween reading and unlinking the label we
restore the last checkpoint from WAL which can need additional WAL.
To fix simply also allow starting when the control file indicates
"shutdown in recovery". There's nicer fixes imaginable, but they'd be
more invasive.
Backpatch to 9.2 where support for taking basebackups from standbys
was added.
Rename parameter action_at_recovery_target to
recovery_target_action suggested by Christoph Berg.
Place into recovery.conf suggested by Fujii Masao,
replacing (deprecating) earlier parameters, per
Michael Paquier.
Transactions can now set their commit timestamp directly as they commit,
or an external transaction commit timestamp can be fed from an outside
system using the new function TransactionTreeSetCommitTsData(). This
data is crash-safe, and truncated at Xid freeze point, same as pg_clog.
This module is disabled by default because it causes a performance hit,
but can be enabled in postgresql.conf requiring only a server restart.
A new test in src/test/modules is included.
Catalog version bumped due to the new subdirectory within PGDATA and a
couple of new SQL functions.
Authors: Álvaro Herrera and Petr Jelínek
Reviewed to varying degrees by Michael Paquier, Andres Freund, Robert
Haas, Amit Kapila, Fujii Masao, Jaime Casanova, Simon Riggs, Steven
Singer, Peter Eisentraut
InitXLogInsert() cannot be called in a critical section, because it
allocates memory. But CreateCheckPoint() did that, when called for the
end-of-recovery checkpoint by the startup process.
In the passing, fix the scratch space allocation in InitXLogInsert to go to
the right memory context. Also update the comment at InitXLOGAccess, which
hasn't been totally accurate since hot standby was introduced (in a hot
standby backend, InitXLOGAccess isn't called at backend startup).
Reported by Michael Paquier
Add a new XLOG_FPI_FOR_HINT record type, and use that for full-page images
generated for hint bit updates, when checksums are enabled. The new record
type is replayed exactly the same as XLOG_FPI, but allows them to be tallied
separately e.g. in pg_xlogdump.
Each WAL record now carries information about the modified relation and
block(s) in a standardized format. That makes it easier to write tools that
need that information, like pg_rewind, prefetching the blocks to speed up
recovery, etc.
There's a whole new API for building WAL records, replacing the XLogRecData
chains used previously. The new API consists of XLogRegister* functions,
which are called for each buffer and chunk of data that is added to the
record. The new API also gives more control over when a full-page image is
written, by passing flags to the XLogRegisterBuffer function.
This also simplifies the XLogReadBufferForRedo() calls. The function can dig
the relation and block number from the WAL record, so they no longer need to
be passed as arguments.
For the convenience of redo routines, XLogReader now disects each WAL record
after reading it, copying the main data part and the per-block data into
MAXALIGNed buffers. The data chunks are not aligned within the WAL record,
but the redo routines can assume that the pointers returned by XLogRecGet*
functions are. Redo routines are now passed the XLogReaderState, which
contains the record in the already-disected format, instead of the plain
XLogRecord.
The new record format also makes the fixed size XLogRecord header smaller,
by removing the xl_len field. The length of the "main data" portion is now
stored at the end of the WAL record, and there's a separate header after
XLogRecord for it. The alignment padding at the end of XLogRecord is also
removed. This compansates for the fact that the new format would otherwise
be more bulky than the old format.
Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera,
Fujii Masao.
Unlogged relations are reset at the end of crash recovery as they're
only synced to disk during a proper shutdown. Unfortunately that and
later steps can fail, e.g. due to running out of space. This reset
was, up to now performed after marking the database as having finished
crash recovery successfully. As out of space errors trigger a crash
restart that could lead to the situation that not all unlogged
relations are reset.
Once that happend usage of unlogged relations could yield errors like
"could not open file "...": No such file or directory". Luckily
clusters that show the problem can be fixed by performing a immediate
shutdown, and starting the database again.
To fix, just call ResetUnloggedRelations(UNLOGGED_RELATION_INIT)
earlier, before marking the database as having successfully recovered.
Discussion: 20140912112246.GA4984@alap3.anarazel.de
Backpatch to 9.1 where unlogged tables were introduced.
Abhijit Menon-Sen and Andres Freund
Now that the backup blocks are appended to the WAL record in xloginsert.c,
XLogInsert doesn't see them anymore and cannot remove them from the version
reconstructed for xlog_outdesc. This makes running with wal_debug=on more
expensive, as we now make (unnecessary) temporary copies of the backup
blocks, but it doesn't seem worth convoluting the code to keep that
optimization.
Reported by Alvaro Herrera.
xlog.c is huge, this makes it a little bit smaller, which is nice. Functions
related to putting together the WAL record are in xloginsert.c, and the
lower level stuff for managing WAL buffers and such are in xlog.c.
Also move the definition of XLogRecord to a separate header file. This
causes churn in the #includes of all the files that write WAL records, and
redo routines, but it avoids pulling in xlog.h into most places.
Reviewed by Michael Paquier, Alvaro Herrera, Andres Freund and Amit Kapila.
The old algorithm was found to not be the usual CRC-32 algorithm, used by
Ethernet et al. We were using a non-reflected lookup table with code meant
for a reflected lookup table. That's a strange combination that AFAICS does
not correspond to any bit-wise CRC calculation, which makes it difficult to
reason about its properties. Although it has worked well in practice, seems
safer to use a well-known algorithm.
Since we're changing the algorithm anyway, we might as well choose a
different polynomial. The Castagnoli polynomial has better error-correcting
properties than the traditional CRC-32 polynomial, even if we had
implemented it correctly. Another reason for picking that is that some new
CPUs have hardware support for calculating CRC-32C, but not CRC-32, let
alone our strange variant of it. This patch doesn't add any support for such
hardware, but a future patch could now do that.
The old algorithm is kept around for tsquery and pg_trgm, which use the
values in indexes that need to remain compatible so that pg_upgrade works.
While we're at it, share the old lookup table for CRC-32 calculation
between hstore, ltree and core. They all use the same table, so might as
well.
Previously the archive recovery always created .ready file for
the last WAL file of the old timeline at the end of recovery even when
it's restored from the archive and has .done file. That is, there was
the case where the WAL file had both .ready and .done files.
This caused the already-archived WAL file to be archived again.
This commit prevents the archive recovery from creating .ready file
for the last WAL file if it has .done file, in order to prevent it from
being archived again.
This bug was added when cascading replication feature was introduced,
i.e., the commit 5286105800.
So, back-patch to 9.2, where cascading replication was added.
Reviewed by Michael Paquier
The duplication originated in cdd46c765, where restartpoints were
introduced.
In LogCheckpointStart's case the duplication actually lead to the
compiler's format string checking not to be effective because the
format string wasn't constant.
Arguably these messages shouldn't be elog(), but ereport() style
messages. That'd even allow to translate the messages... But as
there's more mistakes of that kind in surrounding code, it seems
better to change that separately.
Commit 7dbb606938 added a new CHECKPOINT_FLUSH_ALL flag. As that
commit needed to be backpatched I didn't change the numeric values of
the existing flags as that could lead to nastly problems if any
external code issued checkpoints. That's not a concern on master, so
renumber them there.
Also add a comment about CHECKPOINT_FLUSH_ALL above
CreateCheckPoint().
CREATE DATABASE and ALTER DATABASE .. SET TABLESPACE copy the source
database directory on the filesystem level. To ensure the on disk
state is consistent they block out users of the affected database and
force a checkpoint to flush out all data to disk. Unfortunately, up to
now, that checkpoint didn't flush out dirty buffers from unlogged
relations.
That bug means there could be leftover dirty buffers in either the
template database, or the database in its old location. Leading to
problems when accessing relations in an inconsistent state; and to
possible problems during shutdown in the SET TABLESPACE case because
buffers belonging files that don't exist anymore are flushed.
This was reported in bug #10675 by Maxim Boguk.
Fix by Pavan Deolasee, modified somewhat by me. Reviewed by MauMau and
Fujii Masao.
Backpatch to 9.1 where unlogged tables were introduced.
I left the GUC in place for the beta period, so that people could experiment
with different values. No-one's come up with any data that a different value
would be better under some circumstances, so rather than try to document to
users what the GUC, let's just hard-code the current value, 8.
As noted in http://bugs.debian.org/763098 there is a conflict between
postgres' definition of CACHE_LINE_SIZE and the definition by various
*bsd platforms. It's debatable who has the right to define such a
name, but postgres' use was only introduced in 375d8526f2 (9.4), so
it seems like a good idea to rename it.
Discussion: 20140930195756.GC27407@msg.df7cb.de
Per complaint of Christoph Berg in the above email, although he's not
the original bug reporter.
Backpatch to 9.4 where the define was introduced.
For the reason outlined in df4077cda2 also remove volatile qualifiers
from xlog.c. Some of these uses of volatile have been added after
noticing problems back when spinlocks didn't imply compiler
barriers. So they are a good test - in fact removing the volatiles
breaks when done without the barriers in spinlocks present.
Several uses of volatile remain where they are explicitly used to
access shared memory without locks. These locations are ok with
slightly out of date data, but removing the volatile might lead to the
variables never being reread from memory. These uses could also be
replaced by barriers, but that's a separate change of doubtful value.
There are four weaknesses in728f152e07f998d2cb4fe5f24ec8da2c3bda98f2:
* append_init() in heapdesc.c was ugly and required that rm_identify
return values are only valid till the next call. Instead just add a
couple more switch() cases for the INIT_PAGE cases. Now the returned
value will always be valid.
* a couple rm_identify() callbacks missed masking xl_info with
~XLR_INFO_MASK.
* pg_xlogdump didn't map a NULL rm_identify to UNKNOWN or a similar
string.
* append_init() was called when id=NULL - which should never actually
happen. But it's better to be careful.
This is primarily useful for the upcoming pg_xlogdump --stats feature,
but also allows to remove some duplicated code in the rmgr_desc
routines.
Due to the separation and harmonization, the output of dipsplayed
records changes somewhat. But since this isn't enduser oriented
content that's ok.
It's potentially desirable to further change pg_xlogdump's display of
records. It previously wasn't possible to show the record type
separately from the description forcing it to be in the last
column. But that's better done in a separate commit.
Author: Abhijit Menon-Sen, slightly editorialized by me
Reviewed-By: Álvaro Herrera, Andres Freund, and Heikki Linnakangas
Discussion: 20140604104716.GA3989@toroid.org
log_newpage is used by many indexams, in addition to heap, but for
historical reasons it's always been part of the heapam rmgr. Starting with
9.3, we have another WAL record type for logging an image of a page,
XLOG_FPI. Simplify things by moving log_newpage and log_newpage_buffer to
xlog.c, and switch to using the XLOG_FPI record type.
Bump the WAL version number because the code to replay the old HEAP_NEWPAGE
records is removed.
There were several oversights in recovery code where COMMIT/ABORT PREPARED
records were ignored:
* pg_last_xact_replay_timestamp() (wasn't updated for 2PC commits)
* recovery_min_apply_delay (2PC commits were applied immediately)
* recovery_target_xid (recovery would not stop if the XID used 2PC)
The first of those was reported by Sergiy Zuban in bug #11032, analyzed by
Tom Lane and Andres Freund. The bug was always there, but was masked before
commit d19bd29f07, because COMMIT PREPARED
always created an extra regular transaction that was WAL-logged.
Backpatch to all supported versions (older versions didn't have all the
features and therefore didn't have all of the above bugs).
Nothing in the checkpointer calls InitXLOGAccess(), so WALInsertLocks
never got initialized there. Without EXEC_BACKEND, it works anyway
because the correct value is inherited from the postmaster, but
with EXEC_BACKEND we've got a problem. The problem appears to have
been introduced by commit 68a2e52bba.
To fix, move the relevant initialization steps from InitXLOGAccess()
to XLOGShmemInit(), making this more parallel to what we do
elsewhere.
Amit Kapila
The assertion failed if WAL_DEBUG or LWLOCK_STATS was enabled; fix that by
using separate memory contexts for the allocations made within those code
blocks.
This patch introduces a mechanism for marking any memory context as allowed
in a critical section. Previously ErrorContext was exempt as a special case.
Instead of a blanket exception of the checkpointer process, only exempt the
memory context used for the pending ops hash table.
Instead of truncating pg_multixact at vacuum time, do it only at
checkpoint time. The reason for doing it this way is twofold: first, we
want it to delete only segments that we're certain will not be required
if there's a crash immediately after the removal; and second, we want to
do it relatively often so that older files are not left behind if
there's an untimely crash.
Per my proposal in
http://www.postgresql.org/message-id/20140626044519.GJ7340@eldon.alvh.no-ip.org
we now execute the truncation in the checkpointer process rather than as
part of vacuum. Vacuum is in only charge of maintaining in shared
memory the value to which it's possible to truncate the files; that
value is stored as part of checkpoints also, and so upon recovery we can
reuse the same value to re-execute truncate and reset the
oldest-value-still-safe-to-use to one known to remain after truncation.
Per bug reported by Jeff Janes in the course of his tests involving
bug #8673.
While at it, update some comments that hadn't been updated since
multixacts were changed.
Backpatch to 9.3, where persistency of pg_multixact files was
introduced by commit 0ac5ad5134.
Heikki Linnakangas
Fujii Masao
Andres Freund
Alvaro Herrera
Bruce Momjian
Tom Lane
Simon Riggs
Peter Eisentraut
Robert Haas