If VirtualXactLock() has to wait for a transaction that holds its VXID lock
as a fast-path lock, it must first convert the fast-path lock to a regular
lock. It failed to take the required "partition" lock on the main
shared-memory lock table while doing so. This is the direct cause of the
assert failure in GetLockStatusData() recently observed in the buildfarm,
but more worryingly it could result in arbitrary corruption of the shared
lock table if some other process were concurrently engaged in modifying the
same partition of the lock table. Fortunately, VirtualXactLock() is only
used by CREATE INDEX CONCURRENTLY and DROP INDEX CONCURRENTLY, so the
opportunities for failure are fewer than they might have been.
In passing, improve some comments and be a bit more consistent about
order of operations.
In situations where there are over 8MB of empty pages at the end of
a table, the truncation work for trailing empty pages takes longer
than deadlock_timeout, and there is frequent access to the table by
processes other than autovacuum, there was a problem with the
autovacuum worker process being canceled by the deadlock checking
code. The truncation work done by autovacuum up that point was
lost, and the attempt tried again by a later autovacuum worker. The
attempts could continue indefinitely without making progress,
consuming resources and blocking other processes for up to
deadlock_timeout each time.
This patch has the autovacuum worker checking whether it is
blocking any other thread at 20ms intervals. If such a condition
develops, the autovacuum worker will persist the work it has done
so far, release its lock on the table, and sleep in 50ms intervals
for up to 5 seconds, hoping to be able to re-acquire the lock and
try again. If it is unable to get the lock in that time, it moves
on and a worker will try to continue later from the point this one
left off.
While this patch doesn't change the rules about when and what to
truncate, it does cause the truncation to occur sooner, with less
blocking, and with the consumption of fewer resources when there is
contention for the table's lock.
The only user-visible change other than improved performance is
that the table size during truncation may change incrementally
instead of just once.
This problem exists in all supported versions but is infrequently
reported, although some reports of performance problems when
autovacuum runs might be caused by this. Initial commit is just the
master branch, but this should probably be backpatched once the
build farm and general developer usage confirm that there are no
surprising effects.
Jan Wieck
Background workers are postmaster subprocesses that run arbitrary
user-specified code. They can request shared memory access as well as
backend database connections; or they can just use plain libpq frontend
database connections.
Modules listed in shared_preload_libraries can register background
workers in their _PG_init() function; this is early enough that it's not
necessary to provide an extra GUC option, because the necessary extra
resources can be allocated early on. Modules can install more than one
bgworker, if necessary.
Care is taken that these extra processes do not interfere with other
postmaster tasks: only one such process is started on each ServerLoop
iteration. This means a large number of them could be waiting to be
started up and postmaster is still able to quickly service external
connection requests. Also, shutdown sequence should not be impacted by
a worker process that's reasonably well behaved (i.e. promptly responds
to termination signals.)
The current implementation lets worker processes specify their start
time, i.e. at what point in the server startup process they are to be
started: right after postmaster start (in which case they mustn't ask
for shared memory access), when consistent state has been reached
(useful during recovery in a HOT standby server), or when recovery has
terminated (i.e. when normal backends are allowed).
In case of a bgworker crash, actions to take depend on registration
data: if shared memory was requested, then all other connections are
taken down (as well as other bgworkers), just like it were a regular
backend crashing. The bgworker itself is restarted, too, within a
configurable timeframe (which can be configured to be never).
More features to add to this framework can be imagined without much
effort, and have been discussed, but this seems good enough as a useful
unit already.
An elementary sample module is supplied.
Author: Álvaro Herrera
This patch is loosely based on prior patches submitted by KaiGai Kohei,
and unsubmitted code by Simon Riggs.
Reviewed by: KaiGai Kohei, Markus Wanner, Andres Freund,
Heikki Linnakangas, Simon Riggs, Amit Kapila
Rename PGXACT->inCommit flag into delayChkpt flag,
and generalise comments to allow use in other situations,
such as the forthcoming potential use in checksum patch.
Replace wait loop to look for VXIDs with delayChkpt set.
No user visible changes, not behaviour changes at present.
Simon Riggs, reviewed and rebased by Jeff Davis
This reverts commit c11130690d in favor of
actually fixing the problem: namely, that we should never have been
modifying the checkpoint record's nextXid at this point to begin with.
The nextXid should match the state as of the checkpoint's logical WAL
position (ie the redo point), not the state as of its physical position.
It's especially bogus to advance it in some wal_levels and not others.
In any case there is no need for the checkpoint record to carry the
same nextXid shown in the XLOG_RUNNING_XACTS record just emitted by
LogStandbySnapshot, as any replay operation will already have adopted
that value as current.
This fixes bug #7710 from Tarvi Pillessaar, and probably also explains bug
#6291 from Daniel Farina, in that if a checkpoint were in progress at the
instant of XID wraparound, the epoch bump would be lost as reported.
(And, of course, these days there's at least a 50-50 chance of a checkpoint
being in progress at any given instant.)
Diagnosed by me and independently by Andres Freund. Back-patch to all
branches supporting hot standby.
Previously we stored all xids mixed together.
Now we store top-level xids first, followed
by all subxids. Also skip logging any subxids
if the snapshot is suboverflowed, since there
are potentially large numbers of them and they
are not useful in that case anyway. Has value
in the envisaged design for decoding of WAL.
No planned effect on Hot Standby.
Andres Freund, reviewed by me
Files opened with BasicOpenFile or PathNameOpenFile are not automatically
cleaned up on error. That puts unnecessary burden on callers that only want
to keep the file open for a short time. There is AllocateFile, but that
returns a buffered FILE * stream, which in many cases is not the nicest API
to work with. So add function called OpenTransientFile, which returns a
unbuffered fd that's cleaned up like the FILE* returned by AllocateFile().
This plugs a few rare fd leaks in error cases:
1. copy_file() - fixed by by using OpenTransientFile instead of BasicOpenFile
2. XLogFileInit() - fixed by adding close() calls to the error cases. Can't
use OpenTransientFile here because the fd is supposed to persist over
transaction boundaries.
3. lo_import/lo_export - fixed by using OpenTransientFile instead of
PathNameOpenFile.
In addition to plugging those leaks, this replaces many BasicOpenFile() calls
with OpenTransientFile() that were not leaking, because the code meticulously
closed the file on error. That wasn't strictly necessary, but IMHO it's good
for robustness.
The same leaks exist in older versions, but given the rarity of the issues,
I'm not backpatching this. Not yet, anyway - it might be good to backpatch
later, after this mechanism has had some more testing in master branch.
errcontext() is typically used in an error context callback function, not
within an ereport() invocation like e.g errmsg and errdetail are. That means
that the message domain that the TEXTDOMAIN magic in ereport() determines
is not the right one for the errcontext() calls. The message domain needs to
be determined by the C file containing the errcontext() call, not the file
containing the ereport() call.
Fix by turning errcontext() into a macro that passes the TEXTDOMAIN to use
for the errcontext message. "errcontext" was used in a few places as a
variable or struct field name, I had to rename those out of the way, now
that errcontext is a macro.
We've had this problem all along, but this isn't doesn't seem worth
backporting. It's a fairly minor issue, and turning errcontext from a
function to a macro requires at least a recompile of any external code that
calls errcontext().
If an SMgrRelation is not "owned" by a relcache entry, don't allow it to
live past transaction end. This design allows the same SMgrRelation to be
used for blind writes of multiple blocks during a transaction, but ensures
that we don't hold onto such an SMgrRelation indefinitely. Because an
SMgrRelation typically corresponds to open file descriptors at the fd.c
level, leaving it open when there's no corresponding relcache entry can
mean that we prevent the kernel from reclaiming deleted disk space.
(While CacheInvalidateSmgr messages usually fix that, there are cases
where they're not issued, such as DROP DATABASE. We might want to add
some more sinval messaging for that, but I'd be inclined to keep this
type of logic anyway, since allowing VFDs to accumulate indefinitely
for blind-written relations doesn't seem like a good idea.)
This code replaces a previous attempt towards the same goal that proved
to be unreliable. Back-patch to 9.1 where the previous patch was added.
This reverts commit fba105b109.
That approach had problems with the smgr-level state not tracking what
we really want to happen, and with the VFD-level state not tracking the
smgr-level state very well either. In consequence, it was still possible
to hold kernel file descriptors open for long-gone tables (as in recent
report from Tore Halset), and yet there were also cases of FDs being closed
undesirably soon. A replacement implementation will follow.
In the previous coding, new backend processes would attempt to create their
self-pipe during the OwnLatch call in InitProcess. However, pipe creation
could fail if the kernel is short of resources; and the system does not
recover gracefully from a FATAL error right there, since we have armed the
dead-man switch for this process and not yet set up the on_shmem_exit
callback that would disarm it. The postmaster then forces an unnecessary
database-wide crash and restart, as reported by Sean Chittenden.
There are various ways we could rearrange the code to fix this, but the
simplest and sanest seems to be to split out creation of the self-pipe into
a new function InitializeLatchSupport, which must be called from a place
where failure is allowed. For most processes that gets called in
InitProcess or InitAuxiliaryProcess, but processes that don't call either
but still use latches need their own calls.
Back-patch to 9.1, which has only a part of the latch logic that 9.2 and
HEAD have, but nonetheless includes this bug.
Do read/write permissions checks at most once per large object descriptor,
not once per lo_read or lo_write call as before. The repeated tests were
quite useless in the read case since the snapshot-based tests were
guaranteed to produce the same answer every time. In the write case,
the extra tests could in principle detect revocation of write privileges
after a series of writes has started --- but there's a race condition there
anyway, since we'd check privileges before performing and certainly before
committing the write. So there's no real advantage to checking every
single time, and we might as well redefine it as "only check the first
time".
On the same reasoning, remove the LargeObjectExists checks in inv_write
and inv_truncate. We already checked existence when the descriptor was
opened, and checking again doesn't provide any real increment of safety
that would justify the cost.
Fix broken-on-bigendian-machines byte-swapping functions, add missed update
of alternate regression expected file, improve error reporting, remove some
unnecessary code, sync testlo64.c with current testlo.c (it seems to have
been cloned from a very old copy of that), assorted cosmetic improvements.
4TB large objects (standard 8KB BLCKSZ case). For this purpose new
libpq API lo_lseek64, lo_tell64 and lo_truncate64 are added. Also
corresponding new backend functions lo_lseek64, lo_tell64 and
lo_truncate64 are added. inv_api.c is changed to handle 64-bit
offsets.
Patch contributed by Nozomi Anzai (backend side) and Yugo Nagata
(frontend side, docs, regression tests and example program). Reviewed
by Kohei Kaigai. Committed by Tatsuo Ishii with minor editings.
Recovery code documents clearly that a shutdown checkpoint is executed at
end of recovery - a shutdown checkpoint WAL record is written but the buffer
manager had been altered to treat end of recovery as a normal checkpoint.
This bug exacerbates the bufmgr relpersistence bug.
Bug spotted by Andres Freund, patch by me.
This reduces unnecessary exposure of other headers through htup.h, which
is very widely included by many files.
I have chosen to move the function prototypes to the new file as well,
because that means htup.h no longer needs to include tupdesc.h. In
itself this doesn't have much effect in indirect inclusion of tupdesc.h
throughout the tree, because it's also required by execnodes.h; but it's
something to explore in the future, and it seemed best to do the htup.h
change now while I'm busy with it.
The GUC check hooks for transaction_read_only and transaction_isolation
tried to check RecoveryInProgress(), so as to disallow setting read/write
mode or serializable isolation level (respectively) in hot standby
sessions. However, GUC check hooks can be called in many situations where
we're not connected to shared memory at all, resulting in a crash in
RecoveryInProgress(). Among other cases, this results in EXEC_BACKEND
builds crashing during child process start if default_transaction_isolation
is serializable, as reported by Heikki Linnakangas. Protect those calls
by silently allowing any setting when not inside a transaction; which is
okay anyway since these GUCs are always reset at start of transaction.
Also, add a check to GetSerializableTransactionSnapshot() to complain
if we are in hot standby. We need that check despite the one in
check_XactIsoLevel() because default_transaction_isolation could be
serializable. We don't want to complain any sooner than this in such
cases, since that would prevent running transactions at all in such a
state; but a transaction can be run, if SET TRANSACTION ISOLATION is done
before setting a snapshot. Per report some months ago from Robert Haas.
Back-patch to 9.1, since these problems were introduced by the SSI patch.
Kevin Grittner and Tom Lane, with ideas from Heikki Linnakangas
Move discussion of why our algorithm for taking snapshots in recovery
to a more appropriate location in the function, and delete incorrect
mention of taking a lock.
In the original coding of the autovacuum cancel feature, commit
acac68b2bc, an autovacuum process was
considered a target for cancellation if it was found to hard-block any
process examined in the deadlock search. This patch tightens the test so
that the autovacuum must directly hard-block the current process. This
should make the behavior more predictable in general, and in particular
it ensures that an autovacuum will not be canceled with less than
deadlock_timeout grace period. In the old coding, it was possible for an
autovacuum to be canceled almost instantly, given unfortunate timing of two
or more other processes' lock attempts.
This also justifies the logging methodology in the recent commit
d7318d43d891bd63e82dcfc27948113ed7b1db80; without this restriction, that
patch isn't providing enough information to see the connection of the
canceling process to the autovacuum. Like that one, patch all the way
back.
The old message was at DEBUG2, so typically it didn't show up in the
log at all. As a result, in most cases where autovacuum was canceled,
the only information that was logged was the table being vacuumed,
with no indication as to what problem caused the cancel. Crank up
the level to LOG and add some more details to assist with debugging.
Back-patch all the way, per discussion on pgsql-hackers.
We should avoid calling sync_file_range or posix_fadvise in this case,
since (a) we don't really care if the data gets synced, and might as
well save the kernel calls; (b) at least on Linux we know that the
kernel might block us until it's scheduled the write.
Also, avoid making a useless second traversal of the directory tree
if we're not actually going to call fsync(2) after all.
Instead of having one hash table entry per relation/fork/segment, just have
one per relation, and use bitmapsets to represent which specific segments
need to be fsync'd. This eliminates the need to scan the whole hash table
to implement FORGET_RELATION_FSYNC, which fixes the O(N^2) behavior
recently demonstrated by Jeff Janes for cases involving lots of TRUNCATE or
DROP TABLE operations during a single checkpoint cycle. Per an idea from
Robert Haas.
(FORGET_DATABASE_FSYNC still sucks, but since dropping a database is a
pretty expensive operation anyway, we'll live with that.)
In passing, improve the delayed-unlink code: remove the pass over the list
in mdpreckpt, since it wasn't doing anything for us except supporting a
useless Assert in mdpostckpt, and fix mdpostckpt so that it will absorb
fsync requests every so often when clearing a large backlog of deletion
requests.
We were sending one per fork, but a little bit of refactoring allows us
to send just one request with forknum == InvalidForkNumber. This not only
reduces pressure on the shared-memory request queue, but saves repeated
traversals of the checkpointer's hash table.
mdinit() was misusing IsBootstrapProcessingMode() to decide whether to
create an fsync pending-operations table in the current process. This led
to creating a table not only in the startup and checkpointer processes as
intended, but also in the bgwriter process, not to mention other auxiliary
processes such as walwriter and walreceiver. Creation of the table in the
bgwriter is fatal, because it absorbs fsync requests that should have gone
to the checkpointer; instead they just sit in bgwriter local memory and are
never acted on. So writes performed by the bgwriter were not being fsync'd
which could result in data loss after an OS crash. I think there is no
live bug with respect to walwriter and walreceiver because those never
perform any writes of shared buffers; but the potential is there for
future breakage in those processes too.
To fix, make AuxiliaryProcessMain() export the current process's
AuxProcType as a global variable, and then make mdinit() test directly for
the types of aux process that should have a pendingOpsTable. Having done
that, we might as well also get rid of the random bool flags such as
am_walreceiver that some of the aux processes had grown. (Note that we
could not have fixed the bug by examining those variables in mdinit(),
because it's called from BaseInit() which is run by AuxiliaryProcessMain()
before entering any of the process-type-specific code.)
Back-patch to 9.2, where the problem was introduced by the split-up of
bgwriter and checkpointer processes. The bogus pendingOpsTable exists
in walwriter and walreceiver processes in earlier branches, but absent
any evidence that it causes actual problems there, I'll leave the older
branches alone.
In all branches back to 8.3, this patch fixes a questionable assumption in
CompactCheckpointerRequestQueue/CompactBgwriterRequestQueue that there are
no uninitialized pad bytes in the request queue structs. This would only
cause trouble if (a) there were such pad bytes, which could happen in 8.4
and up if the compiler makes enum ForkNumber narrower than 32 bits, but
otherwise would require not-currently-planned changes in the widths of
other typedefs; and (b) the kernel has not uniformly initialized the
contents of shared memory to zeroes. Still, it seems a tad risky, and we
can easily remove any risk by pre-zeroing the request array for ourselves.
In addition to that, we need to establish a coding rule that struct
RelFileNode can't contain any padding bytes, since such structs are copied
into the request array verbatim. (There are other places that are assuming
this anyway, it turns out.)
In 9.1 and up, the risk was a bit larger because we were also effectively
assuming that struct RelFileNodeBackend contained no pad bytes, and with
fields of different types in there, that would be much easier to break.
However, there is no good reason to ever transmit fsync or delete requests
for temp files to the bgwriter/checkpointer, so we can revert the request
structs to plain RelFileNode, getting rid of the padding risk and saving
some marginal number of bytes and cycles in fsync queue manipulation while
we are at it. The savings might be more than marginal during deletion of
a temp relation, because the old code transmitted an entirely useless but
nonetheless expensive-to-process ForgetRelationFsync request to the
background process, and also had the background process perform the file
deletion even though that can safely be done immediately.
In addition, make some cleanup of nearby comments and small improvements to
the code in CompactCheckpointerRequestQueue/CompactBgwriterRequestQueue.
Management of timeouts was getting a little cumbersome; what we
originally had was more than enough back when we were only concerned
about deadlocks and query cancel; however, when we added timeouts for
standby processes, the code got considerably messier. Since there are
plans to add more complex timeouts, this seems a good time to introduce
a central timeout handling module.
External modules register their timeout handlers during process
initialization, and later enable and disable them as they see fit using
a simple API; timeout.c is in charge of keeping track of which timeouts
are in effect at any time, installing a common SIGALRM signal handler,
and calling setitimer() as appropriate to ensure timely firing of
external handlers.
timeout.c additionally supports pluggable modules to add their own
timeouts, though this capability isn't exercised anywhere yet.
Additionally, as of this commit, walsender processes are aware of
timeouts; we had a preexisting bug there that made those ignore SIGALRM,
thus being subject to unhandled deadlocks, particularly during the
authentication phase. This has already been fixed in back branches in
commit 0bf8eb2a, which see for more details.
Main author: Zoltán Böszörményi
Some review and cleanup by Álvaro Herrera
Extensive reworking by Tom Lane
The Solaris Studio compiler warns about these instances, unlike more
mainstream compilers such as gcc. But manual inspection showed that
the code is clearly not reachable, and we hope no worthy compiler will
complain about removing this code.
Historically we have not worried about fsync'ing anything during initdb
(in fact, initdb intentionally passes -F to each backend launch to prevent
it from fsync'ing). But with filesystems getting more aggressive about
caching data, that's not such a good plan anymore. Make initdb do a pass
over the finished data directory tree to fsync everything. For testing
purposes, the -N/--nosync flag can be used to restore the old behavior.
Also, testing shows that on Linux, sync_file_range() is much faster than
posix_fadvise() for hinting to the kernel that an fsync is coming,
apparently because the latter blocks on a rather small request queue while
the former doesn't. So use this function if available in initdb, and also
in the backend's pg_flush_data() (where it currently will affect only the
speed of CREATE DATABASE's cloning step).
We will later make pg_regress invoke initdb with the --nosync flag
to avoid slowing down cases such as "make check" in contrib. But
let's not do so until we've shaken out any portability issues in this
patch.
Jeff Davis, reviewed by Andres Freund
This simplifies code that needs to do arithmetic on XLogRecPtrs.
To avoid changing on-disk format of data pages, the LSN on data pages is
still stored in the old format. That should keep pg_upgrade happy. However,
we have XLogRecPtrs embedded in the control file, and in the structs that
are sent over the replication protocol, so this changes breaks compatibility
of pg_basebackup and server. I didn't do anything about this in this patch,
per discussion on -hackers, the right thing to do would to be to change the
replication protocol to be architecture-independent, so that you could use
a newer version of pg_receivexlog, for example, against an older server
version.
This speeds up reassigning locks to the parent owner, when the transaction
holds a lot of locks, but only a few of them belong to the current resource
owner. This is particularly helps pg_dump when dumping a large number of
objects.
The cache can hold up to 15 locks in each resource owner. After that, the
cache is marked as overflowed, and we fall back to the old method of
scanning the whole local lock table. The tradeoff here is that the cache has
to be scanned whenever a lock is released, so if the cache is too large,
lock release becomes more expensive. 15 seems enough to cover pg_dump, and
doesn't have much impact on lock release.
Jeff Janes, reviewed by Amit Kapila and Heikki Linnakangas.
When HS startup is deferred because of overflowed subtransactions, ensure
that we re-initialize KnownAssignedXids for when both existing and incoming
snapshots have non-zero qualifying xids.
Fixes bug #6661 reported by Valentine Gogichashvili.
Analysis and fix by Andres Freund
This provides a speedup of about 4X when NBuffers is large enough.
There is also a useful reduction in sinval traffic, since we
only do CacheInvalidateSmgr() once not once per fork.
Simon Riggs, reviewed and somewhat revised by Tom Lane
DropRelFileNodeBuffers, DropDatabaseBuffers, FlushRelationBuffers, and
FlushDatabaseBuffers have to scan the whole shared_buffers pool because
we have no index structure that would find the target buffers any more
efficiently than that. This gets expensive with large NBuffers. We can
shave some cycles from these loops by prechecking to see if the current
buffer is interesting before we acquire the buffer header lock.
Ordinarily such a test would be unsafe, but in these cases it should be
safe because we are already assuming that the caller holds a lock that
prevents any new target pages from being loaded into the buffer pool
concurrently. Therefore, no buffer tag should be changing to a value of
interest, only away from a value of interest. So a false negative match
is impossible, while a false positive is safe because we'll recheck after
acquiring the buffer lock. Initial testing says that this speeds these
loops by a factor of 2X to 3X on common Intel hardware.
Patch for DropRelFileNodeBuffers by Jeff Janes (based on an idea of
Heikki's); extended to the remaining sequential scans by Tom Lane
First, the previous code failed to account for the fact that, during Hot
Standby operation, the startup process takes AccessExclusiveLocks on
relations without setting MyDatabaseId. This resulted in fast path
strong lock counts failing to be incremented with the startup process
took locks, which in turn allowed conflicting lock requests to succeed
when they should not have. Report by Erik Rijkers, diagnosis by Heikki
Linnakangas.
Second, LockReleaseAll() failed to honor the allLocks and lockmethodid
restrictions with respect to fast-path locks. It's not clear to me
whether this produces any user-visible breakage at the moment, but it's
certainly wrong. Rearrange order of operations in LockReleaseAll to fix.
Noted by Tom Lane.
In commit d526575f89, we changed things so
that buffer usage counts are incremented when the buffer is pinned, rather
than when it is unpinned, but the README file didn't get the memo.
Report by Amit Kapila.
When the "hot" members of PGPROC were split off to separate PGXACT structs,
many PGPROC fields referred to in comments were moved to PGXACT, but the
comments were neglected in the commit. Mostly this is just a search/replace
of PGPROC with PGXACT, but the way the dummy PGPROC entries are created for
prepared transactions changed more, making some of the comments totally
bogus.
Noah Misch
