array_base and array_stride were added so that we could identify the
offset of an LWLock within a tranche, but this facility is only very
marginally used apart from the main tranche. So, give every lock in
the main tranche its own tranche ID and get rid of array_base,
array_stride, and all that's attached. For debugging facilities
(Trace_lwlocks and LWLOCK_STATS) print the pointer address of the
LWLock using %p instead of the offset. This is arguably more useful,
and certainly a lot cheaper. Drop the offset-within-tranche from
the information reported to dtrace and from one can't-happen message
inside lwlock.c.
The main user-visible impact of this change is that pg_stat_activity
will now report all waits for LWLocks as "LWLock" rather than
reporting some as "LWLockTranche" and others as "LWLockNamed".
The main motivation for this change is that the need to specify an
array_base and an array_stride is awkward for parallel query. There
is only a very limited supply of tranche IDs so we can't just keep
allocating new ones, and if we try to use the same tranche IDs every
time then we run into trouble when multiple parallel contexts are
use simultaneously. So if we didn't get rid of this mechanism we'd
have to make it even more complicated. By simplifying it in this
way, we instead reduce the size of the generated code for lwlock.c
by about 5%.
Discussion: http://postgr.es/m/CA+TgmoYsFn6NUW1x0AZtupJGUAs1UDY4dJtCN47_Q6D0sP80PA@mail.gmail.com
Previously, the "sem" field of PGPROC varied in size depending on which
kernel semaphore API we were using. That was okay as long as there was
only one likely choice per platform, but in the wake of commit ecb0d20a9,
that assumption seems rather shaky. It doesn't seem out of the question
anymore that an extension compiled against one API choice might be loaded
into a postmaster built with another choice. Moreover, this prevents any
possibility of selecting the semaphore API at postmaster startup, which
might be something we want to do in future.
Hence, change PGPROC.sem to be PGSemaphore (i.e. a pointer) for all Unix
semaphore APIs, and turn the pointed-to data into an opaque struct whose
contents are only known within the responsible modules.
For the SysV and unnamed-POSIX APIs, the pointed-to data has to be
allocated elsewhere in shared memory, which takes a little bit of
rejiggering of the InitShmemAllocation code sequence. (I invented a
ShmemAllocUnlocked() function to make that a little cleaner than it used
to be. That function is not meant for any uses other than the ones it
has now, but it beats having InitShmemAllocation() know explicitly about
allocation of space for semaphores and spinlocks.) This change means an
extra indirection to access the semaphore data, but since we only touch
that when blocking or awakening a process, there shouldn't be any
meaningful performance penalty. Moreover, at least for the unnamed-POSIX
case on Linux, the sem_t type is quite a bit wider than a pointer, so this
reduces sizeof(PGPROC) which seems like a good thing.
For the named-POSIX API, there's effectively no change: the PGPROC.sem
field was and still is a pointer to something returned by sem_open() in
the postmaster's memory space. Document and check the pre-existing
limitation that this case can't work in EXEC_BACKEND mode.
It did not seem worth unifying the Windows semaphore ABI with the Unix
cases, since there's no likelihood of needing ABI compatibility much less
runtime switching across those cases. However, we can simplify the Windows
code a bit if we define PGSemaphore as being directly a HANDLE, rather than
pointer to HANDLE, so let's do that while we're here. (This also ends up
being no change in what's physically stored in PGPROC.sem. We're just
moving the HANDLE fetch from callees to callers.)
It would take a bunch of additional code shuffling to get to the point of
actually choosing a semaphore API at postmaster start, but the effects
of that would now be localized in the port/XXX_sema.c files, so it seems
like fit material for a separate patch. The need for it is unproven as
yet, anyhow, whereas the ABI risk to extensions seems real enough.
Discussion: https://postgr.es/m/4029.1481413370@sss.pgh.pa.us
Instead of confusingly stating platform-dependent defaults for these
parameters in the comments in postgresql.conf.sample (with the main
entry being a lie on Linux), teach initdb to install the correct
platform-dependent value in postgresql.conf, similarly to the way
we handle other platform-dependent defaults. This won't do anything
for existing 9.6 installations, but since it's effectively only a
documentation improvement, that seems OK.
Since this requires initdb to have access to the default values,
move the #define's for those to pg_config_manual.h; the original
placement in bufmgr.h is unworkable because that file can't be
included by frontend programs.
Adjust the default value for wal_writer_flush_after so that it is 1MB
regardless of XLOG_BLCKSZ, conforming to what is stated in both the
SGML docs and postgresql.conf. (We could alternatively make it scale
with XLOG_BLCKSZ, but I'm not sure I see the point.)
Copy-edit related SGML documentation.
Fabien Coelho and Tom Lane, per a gripe from Tomas Vondra.
Discussion: <30ebc6e3-8358-09cf-44a8-578252938424@2ndquadrant.com>
A new thing also called a "barrier" is proposed, but whether we decide
to take that patch or not, this file seems to have outlived its
usefulness.
Thomas Munro
Condition variables provide a flexible way to sleep until a
cooperating process causes an arbitrary condition to become true. In
simple cases, this can be accomplished with a WaitLatch/ResetLatch
loop; the cooperating process can call SetLatch after performing work
that might cause the condition to be satisfied, and the waiting
process can recheck the condition each time. However, if the process
performing the work doesn't have an easy way to identify which
processes might be waiting, this doesn't work, because it can't
identify which latches to set. Condition variables solve that problem
by internally maintaining a list of waiters; a process that may have
caused some waiter's condition to be satisfied must "signal" or
"broadcast" on the condition variable.
Robert Haas and Thomas Munro
Previously, the handle for the control segment could not be zero, but
some other DSM segment could potentially have a handle value of zero.
However, that means that if someone wanted to store a dsm_handle that
might or might not be valid, they would need a separate boolean to
keep track of whether the associated value is legal. That's annoying,
so change things so that no DSM segment can ever have a handle of 0 -
or as we call it here, DSM_HANDLE_INVALID.
Thomas Munro. This was submitted as part of a much larger patch to
add an malloc-like allocator for dynamic shared memory, but this part
seems like a good idea independently of the rest of the patch.
LockBufferForCleanup() acquires a cleanup lock unconditionally, and
ConditionalLockBufferForCleanup() acquires a cleanup lock if it is
possible to do so without waiting; this patch adds a new API,
IsBufferCleanupOK(), which tests whether an exclusive lock already
held happens to be a cleanup lock. This is possible because a cleanup
lock simply means an exclusive lock plus the assurance any other pins
on the buffer are newer than our own pin. Therefore, just as the
existing functions decide that the exclusive lock that they've just
taken is a cleanup lock if they observe the pin count to be 1, this
new function allows us to observe that the pin count is 1 on a buffer
we've already locked.
This is useful in situations where a backend definitely wishes to
modify the buffer and also wishes to perform cleanup operations if
possible. The patch to eliminate heavyweight locking by hash indexes
uses this, and it may have other applications as well.
Amit Kapila, per a suggestion from me. Some comment adjustments by me
as well.
SCO OpenServer and SCO UnixWare are more or less dead platforms.
We have never had a buildfarm member testing the "sco" port, and
the last "unixware" member was last heard from in 2012, so it's
fair to doubt that the code even compiles anymore on either one.
Remove both ports. We can always undo this if someone shows up
with an interest in maintaining and testing these platforms.
Discussion: <17177.1476136994@sss.pgh.pa.us>
I somehow had assumed that in the spinlock (in turn possibly using
semaphores) based fallback atomics implementation 32 bit writes could be
done without a lock. As far as the write goes that's correct, since
postgres supports only platforms with single-copy atomicity for aligned
32bit writes. But writing without holding the spinlock breaks
read-modify-write operations like pg_atomic_compare_exchange_u32(),
since they'll potentially "miss" a concurrent write, which can't happen
in actual hardware implementations.
In 9.6+ when using the fallback atomics implementation this could lead
to buffer header locks not being properly marked as released, and
potentially some related state corruption. I don't see a related danger
in 9.5 (earliest release with the API), because pg_atomic_write_u32()
wasn't used in a concurrent manner there.
The state variable of local buffers, before this change, were
manipulated using pg_atomic_write_u32(), to avoid unnecessary
synchronization overhead. As that'd not be the case anymore, introduce
and use pg_atomic_unlocked_write_u32(), which does not correctly
interact with RMW operations.
This bug only caused issues when postgres is compiled on platforms
without atomics support (i.e. no common new platform), or when compiled
with --disable-atomics, which explains why this wasn't noticed in
testing.
Reported-By: Tom Lane
Discussion: <14947.1475690465@sss.pgh.pa.us>
Backpatch: 9.5-, where the atomic operations API was introduced.
WaitLatch, WaitLatchOrSocket, and WaitEventSetWait now taken an
additional wait_event_info parameter; legal values are defined in
pgstat.h. This makes it possible to uniquely identify every point in
the core code where we are waiting for a latch; extensions can pass
WAIT_EXTENSION.
Because latches were the major wait primitive not previously covered
by this patch, it is now possible to see information in
pg_stat_activity on a large number of important wait events not
previously addressed, such as ClientRead, ClientWrite, and SyncRep.
Unfortunately, many of the wait events added by this patch will fail
to appear in pg_stat_activity because they're only used in background
processes which don't currently appear in pg_stat_activity. We should
fix this either by creating a separate view for such information, or
else by deciding to include them in pg_stat_activity after all.
Michael Paquier and Robert Haas, reviewed by Alexander Korotkov and
Thomas Munro.
Past refactorings have removed all but one reference to SizeOfIptrData
(and that one place was in a pretty noncritical spot). Since nobody's
complained, it seems probable that there are no supported compilers
that don't think sizeof(ItemPointerData) is 6. If there are, we're
wasting MAXALIGN per heap tuple anyway, so it's rather silly to worry
about whether we can shave space in places like WAL records.
Pavan Deolasee
Discussion: <CABOikdOOawDda4hwLOT6zdA6MFfPLu3Z2YBZkX0JdayNS6JOeQ@mail.gmail.com>
The full generality of deleting an arbitrary number of tuples is no longer
needed, so let's save some code and cycles by replacing the original coding
with an implementation based on PageIndexTupleDelete.
We can always get back the old code from git if we need it again for new
callers (though I don't care for its willingness to mess with line pointers
it wasn't told to mess with).
Discussion: <552.1473445163@sss.pgh.pa.us>
Nowadays this is just a backwards-compatibility wrapper around
PageAddItemExtended, so let's avoid the extra level of function call.
In addition, because pretty much all callers are passing constants
for the two bool arguments, compilers will be able to constant-fold
the conversion to a flags bitmask.
Discussion: <552.1473445163@sss.pgh.pa.us>
PageIndexTupleOverwrite performs approximately the same function as
PageIndexTupleDelete (or PageIndexDeleteNoCompact) followed by PageAddItem
targeting the same item pointer offset. But in the case where the new
tuple is the same size as the old, it avoids shuffling other data around on
the page, because the new tuple is placed where the old one was rather than
being appended to the end of the page. This has been shown to provide a
substantial speedup for some GiST use-cases.
Also, this change allows some API simplifications: we can get rid of
the rather klugy and error-prone PAI_ALLOW_FAR_OFFSET flag for
PageAddItemExtended, since that was used only to cover a corner case
for BRIN that's better expressed by using PageIndexTupleOverwrite.
Note that this patch causes a rather subtle WAL incompatibility: the
physical page content change represented by certain WAL records is now
different than it was before, because while the tuples have the same
itempointer line numbers, the tuples themselves are in different places.
I have not bumped the WAL version number because I think it doesn't matter
unless you are trying to do bitwise comparisons of original and replayed
pages, and in any case we're early in a devel cycle and there will probably
be more WAL changes before v10 gets out the door.
There is probably room to make use of PageIndexTupleOverwrite in SP-GiST
and GIN too, but that is left for a future patch.
Andrey Borodin, reviewed by Anastasia Lubennikova, whacked around a bit
by me
Discussion: <CAJEAwVGQjGGOj6mMSgMwGvtFd5Kwe6VFAxY=uEPZWMDjzbn4VQ@mail.gmail.com>
So far md.c used a linked list of segments. That proved to be a problem
when processing large relations, because every smgr.c/md.c level access
to a page incurred walking through a linked list of all preceding
segments. Thus making accessing pages O(#segments).
Replace the linked list of segments hanging off SMgrRelationData with an
array of opened segments. That allows O(1) access to individual
segments, if they've previously been opened.
Discussion: <20140331101001.GE13135@alap3.anarazel.de>
Reviewed-By: Peter Geoghegan, Tom Lane (in an older version)
A majority of callers seem to have believed that this was the API spec
already, because they omitted any check for a NULL result, and hence
would crash on an out-of-shared-memory failure. The original proposal
was to just add such error checks everywhere, but that does nothing to
prevent similar omissions in future. Instead, let's make ShmemAlloc()
throw the error (so we can remove the caller-side checks that do exist),
and introduce a new function ShmemAllocNoError() that has the previous
behavior of returning NULL, for the small number of callers that need
that and are prepared to do the right thing. This also lets us remove
the rather wishy-washy behavior of printing a WARNING for out-of-shmem,
which never made much sense: either the caller has a strategy for
dealing with that, or it doesn't. It's not ShmemAlloc's business to
decide whether a warning is appropriate.
The v10 release notes will need to call this out as a significant
source-code change. It's likely that it will be a bug fix for
extension callers too, but if not, they'll need to change to using
ShmemAllocNoError().
This is nominally a bug fix, but the odds that it's fixing any live
bug are actually rather small, because in general the requests
being made by the unchecked callers were already accounted for in
determining the overall shmem size, so really they ought not fail.
Between that and the possible impact on extensions, no back-patch.
Discussion: <24843.1472563085@sss.pgh.pa.us>
If you have previously pinned a segment and decide that you don't
actually want to keep it around until shutdown, this new API lets you
remove the pin. This is pretty trivial except on Windows, where it
requires closing the duplicate handle that was used to implement the
pin.
Thomas Munro and Amit Kapila, reviewed by Amit Kapila and by me.
Prior to commit 7882c3b0b9, it was
possible to use LWLocks within DSM segments, but that commit broke
this use case by switching from a doubly linked list to a circular
linked list. Switch back, using a new bit of general infrastructure
for maintaining lists of PGPROCs.
Thomas Munro, reviewed by me.
Previously, we tested for MVCC snapshots to see whether they were too
old, but not TOAST snapshots, which can lead to complaints about missing
TOAST chunks if those chunks are subject to early pruning. Ideally,
the threshold lsn and timestamp for a TOAST snapshot would be that of
the corresponding MVCC snapshot, but since we have no way of deciding
which MVCC snapshot was used to fetch the TOAST pointer, use the oldest
active or registered snapshot instead.
Reported by Andres Freund, who also sketched out what the fix should
look like. Patch by me, reviewed by Amit Kapila.
This coding pattern creates a race condition, because if an interesting
interrupt happens after we've checked InterruptPending but before we reset
our latch, the latch-setting done by the signal handler would get lost,
and then we might block at WaitLatch in the next iteration without ever
noticing the interrupt condition. You can put the CHECK_FOR_INTERRUPTS
before WaitLatch or after ResetLatch, but not between them.
Aside from fixing the bugs, add some explanatory comments to latch.h
to perhaps forestall the next person from making the same mistake.
In HEAD, also replace gather_readnext's direct call of
HandleParallelMessages with CHECK_FOR_INTERRUPTS. It does not seem clean
or useful for this one caller to bypass ProcessInterrupts and go straight
to HandleParallelMessages; not least because that fails to consider the
InterruptPending flag, resulting in useless work both here
(if InterruptPending isn't set) and in the next CHECK_FOR_INTERRUPTS call
(if it is).
This thinko seems to have been introduced in the initial coding of
storage/ipc/shm_mq.c (commit ec9037df2), and then blindly copied into all
the subsequent parallel-query support logic. Back-patch relevant hunks
to 9.4 to extirpate the error everywhere.
Discussion: <1661.1469996911@sss.pgh.pa.us>
While beneficial, both for throughput and average/worst case latency, in
a significant number of workloads, there are other workloads in which
backend_flush_after can cause significant performance regressions in
comparison to < 9.6 releases. The regression is most likely when the hot
data set is bigger than shared buffers, but significantly smaller than
the operating system's page cache.
I personally think that the benefit of enabling backend flush control is
considerably bigger than the potential downsides, but a fair argument
can be made that not regressing is more important than improving
performance/latency. As the latter is the consensus, change the default
to 0.
The other settings introduced in 428b1d6b2 do not have the same
potential for regressions, so leave them enabled.
Benchmarks leading up to changing the default have been performed by
Mithun Cy, Ashutosh Sharma and Robert Haas.
Discussion: CAD__OuhPmc6XH=wYRm_+Q657yQE88DakN4=Ybh2oveFasHkoeA@mail.gmail.com
Transmit the leader's temp-namespace state to workers. This is important
because without it, the workers do not really have the same search path
as the leader. For example, there is no good reason (and no extant code
either) to prevent a worker from executing a temp function that the
leader created previously; but as things stood it would fail to find the
temp function, and then either fail or execute the wrong function entirely.
We still prohibit a worker from creating a temp namespace on its own.
In effect, a worker can only see the session's temp namespace if the leader
had created it before starting the worker, which seems like the right
semantics.
Also, transmit the leader's BackendId to workers, and arrange for workers
to use that when determining the physical file path of a temp relation
belonging to their session. While the original intent was to prevent such
accesses entirely, there were a number of holes in that, notably in places
like dbsize.c which assume they can safely access temp rels of other
sessions anyway. We might as well get this right, as a small down payment
on someday allowing workers to access the leader's temp tables. (With
this change, directly using "MyBackendId" as a relation or buffer backend
ID is deprecated; you should use BackendIdForTempRelations() instead.
I left a couple of such uses alone though, as they're not going to be
reachable in parallel workers until we do something about localbuf.c.)
Move the thou-shalt-not-access-thy-leader's-temp-tables prohibition down
into localbuf.c, which is where it actually matters, instead of having it
in relation_open(). This amounts to recognizing that access to temp
tables' catalog entries is perfectly safe in a worker, it's only the data
in local buffers that is problematic.
Having done all that, we can get rid of the test in has_parallel_hazard()
that says that use of a temp table's rowtype is unsafe in parallel workers.
That test was unduly expensive, and if we really did need such a
prohibition, that was not even close to being a bulletproof guard for it.
(For example, any user-defined function executed in a parallel worker
might have attempted such access.)
Mostly these are just comments but there are a few in documentation
and a handful in code and tests. Hopefully this doesn't cause too much
unnecessary pain for backpatching. I relented from some of the most
common like "thru" for that reason. The rest don't seem numerous
enough to cause problems.
Thanks to Kevin Lyda's tool https://pypi.python.org/pypi/misspellings
This bug appears to have been introduced late in the development of
48354581a4 ("Allow Pin/UnpinBuffer to operate in a lockfree
manner.").
Found while debugging a bug which turned out to be independent of the
commit mentioned above.
Backpatch: -
BRIN was relying on the ability to remove a tuple from an index page,
then putting another tuple in the same line pointer. But PageAddItem
refuses to add a tuple beyond the first free item past the last used
item, and in particular, it rejects an attempt to add an item to an
empty page anywhere other than the first line pointer. PageAddItem
issues a WARNING and indicates to the caller that it failed, which in
turn causes the BRIN calling code to issue a PANIC, so the whole
sequence looks like this:
WARNING: specified item offset is too large
PANIC: failed to add BRIN tuple
To fix, create a new function PageAddItemExtended which is like
PageAddItem except that the two boolean arguments become a flags bitmap;
the "overwrite" and "is_heap" boolean flags in PageAddItem become
PAI_OVERWITE and PAI_IS_HEAP flags in the new function, and a new flag
PAI_ALLOW_FAR_OFFSET enables the behavior required by BRIN.
PageAddItem() retains its original signature, for compatibility with
third-party modules (other callers in core code are not modified,
either).
Also, in the belt-and-suspenders spirit, I added a new sanity check in
brinGetTupleForHeapBlock to raise an error if an TID found in the revmap
is not marked as live by the page header. This causes it to react with
"ERROR: corrupted BRIN index" to the bug at hand, rather than a hard
crash.
Backpatch to 9.5.
Bug reported by Andreas Seltenreich as detected by his handy sqlsmith
fuzzer.
Discussion: https://www.postgresql.org/message-id/87mvni77jh.fsf@elite.ansel.ydns.eu
So far, when a transaction with pending invalidations, but without an
assigned xid, committed, we simply ignored those invalidation
messages. That's problematic, because those are actually sent for a
reason.
Known symptoms of this include that existing sessions on a hot-standby
replica sometimes fail to notice new concurrently built indexes and
visibility map updates.
The solution is to WAL log such invalidations in transactions without an
xid. We considered to alternatively force-assign an xid, but that'd be
problematic for vacuum, which might be run in systems with few xids.
Important: This adds a new WAL record, but as the patch has to be
back-patched, we can't bump the WAL page magic. This means that standbys
have to be updated before primaries; otherwise
"PANIC: standby_redo: unknown op code 32" errors can be encountered.
XXX:
Reported-By: Васильев Дмитрий, Masahiko Sawada
Discussion:
CAB-SwXY6oH=9twBkXJtgR4UC1NqT-vpYAtxCseME62ADwyK5OA@mail.gmail.comCAD21AoDpZ6Xjg=gFrGPnSn4oTRRcwK1EBrWCq9OqOHuAcMMC=w@mail.gmail.com
Even with old_snapshot_threshold = -1 (which disables the "snapshot
too old" feature), performance regressions were seen at moderate to
high concurrency. For example, a one-socket, four-core system
running 200 connections at saturation could see up to a 2.3%
regression, with larger regressions possible on NUMA machines.
By inlining the early (smaller, faster) tests in the
TestForOldSnapshot() function, the i7 case dropped to a 0.2%
regression, which could easily just be noise, and is clearly an
improvement. Further testing will show whether more is needed.
The reverted changes were intended to force a choice of whether any
newly-added BufferGetPage() calls needed to be accompanied by a
test of the snapshot age, to support the "snapshot too old"
feature. Such an accompanying test is needed in about 7% of the
cases, where the page is being used as part of a scan rather than
positioning for other purposes (such as DML or vacuuming). The
additional effort required for back-patching, and the doubt whether
the intended benefit would really be there, have indicated it is
best just to rely on developers to do the right thing based on
comments and existing usage, as we do with many other conventions.
This change should have little or no effect on generated executable
code.
Motivated by the back-patching pain of Tom Lane and Robert Haas
pg_xlogdump includes bufmgr.h. With a compiler that emits code for
static inline functions even when they're unreferenced, that leads
to unresolved external references in the new static-inline version
of BufferGetPage(). So hide it with #ifndef FRONTEND, as we've done
for similar issues elsewhere. Per buildfarm member pademelon.
My previous attempt at doing so, in 80abbeba23, was not sufficient. While that
fixed the problem for bufmgr.c and lwlock.c , s_lock.c still has non-constant
expressions in the struct initializer, because the file/line/function
information comes from the caller of s_lock().
Give up on using a macro, and use a static inline instead.
Discussion: 4369.1460435533@sss.pgh.pa.us
The current definition of init_spin_delay (introduced recently in
48354581a) wasn't C89 compliant. It's not legal to refer to refer to
non-constant expressions, and the ptr argument was one. This, as
reported by Tom, lead to a failure on buildfarm animal pademelon.
The pointer, especially on system systems with ASLR, isn't super helpful
anyway, though. So instead of making init_spin_delay into an inline
function, make s_lock_stuck() report the function name in addition to
file:line and change init_spin_delay() accordingly. While not a direct
replacement, the function name is likely more useful anyway (line
numbers are often hard to interpret in third party reports).
This also fixes what file/line number is reported for waits via
s_lock().
As PG_FUNCNAME_MACRO is now used outside of elog.h, move it to c.h.
Reported-By: Tom Lane
Discussion: 4369.1460435533@sss.pgh.pa.us
The recent patch to make Pin/UnpinBuffer lockfree in the hot
path (48354581a), accidentally used pg_atomic_fetch_or_u32() in
MarkLocalBufferDirty(). Other code operating on local buffers was
careful to only use pg_atomic_read/write_u32 which just read/write from
memory; to avoid unnecessary overhead.
On its own that'd just make MarkLocalBufferDirty() slightly less
efficient, but in addition InitLocalBuffers() doesn't call
pg_atomic_init_u32() - thus the spinlock fallback for the atomic
operations isn't initialized. That in turn caused, as reported by Tom,
buildfarm animal gaur to fail. As those errors are actually useful
against this type of error, continue to omit - intentionally this time -
initialization of the atomic variable.
In addition, add an explicit note about only using pg_atomic_read/write
on local buffers's state to BufferDesc's description.
Reported-By: Tom Lane
Discussion: 1881.1460431476@sss.pgh.pa.us
It's silly to define these counts as narrower than they might someday
need to be. Also, I believe that the BLCKSZ * nflush calculation in
mdwriteback was capable of overflowing an int.
I was initially concerned that the some of the hundreds of
references to BufferGetPage() where the literal
BGP_NO_SNAPSHOT_TEST were passed might not optimize as well as a
macro, leading to some hard-to-find performance regressions in
corner cases. Inspection of disassembled code has shown identical
code at all inspected locations, and the size difference doesn't
amount to even one byte per such call. So make it readable.
Per gripes from Álvaro Herrera and Tom Lane
Previously we used a spinlock, in adition to the atomically manipulated
->state field, to protect the wait queue. But it's pretty simple to
instead perform the locking using a flag in state.
Due to 6150a1b0 BufferDescs, on platforms (like PPC) with > 1 byte
spinlocks, increased their size above 64byte. As 64 bytes are the size
we pad allocated BufferDescs to, this can increase false sharing;
causing performance problems in turn. Together with the previous commit
this reduces the size to <= 64 bytes on all common platforms.
Author: Andres Freund
Discussion: CAA4eK1+ZeB8PMwwktf+3bRS0Pt4Ux6Rs6Aom0uip8c6shJWmyg@mail.gmail.com20160327121858.zrmrjegmji2ymnvr@alap3.anarazel.de
Pinning/Unpinning a buffer is a very frequent operation; especially in
read-mostly cache resident workloads. Benchmarking shows that in various
scenarios the spinlock protecting a buffer header's state becomes a
significant bottleneck. The problem can be reproduced with pgbench -S on
larger machines, but can be considerably worse for queries which touch
the same buffers over and over at a high frequency (e.g. nested loops
over a small inner table).
To allow atomic operations to be used, cram BufferDesc's flags,
usage_count, buf_hdr_lock, refcount into a single 32bit atomic variable;
that allows to manipulate them together using 32bit compare-and-swap
operations. This requires reducing MAX_BACKENDS to 2^18-1 (which could
be lifted by using a 64bit field, but it's not a realistic configuration
atm).
As not all operations can easily implemented in a lockfree manner,
implement the previous buf_hdr_lock via a flag bit in the atomic
variable. That way we can continue to lock the header in places where
it's needed, but can get away without acquiring it in the more frequent
hot-paths. There's some additional operations which can be done without
the lock, but aren't in this patch; but the most important places are
covered.
As bufmgr.c now essentially re-implements spinlocks, abstract the delay
logic from s_lock.c into something more generic. It now has already two
users, and more are coming up; there's a follupw patch for lwlock.c at
least.
This patch is based on a proof-of-concept written by me, which Alexander
Korotkov made into a fully working patch; the committed version is again
revised by me. Benchmarking and testing has, amongst others, been
provided by Dilip Kumar, Alexander Korotkov, Robert Haas.
On a large x86 system improvements for readonly pgbench, with a high
client count, of a factor of 8 have been observed.
Author: Alexander Korotkov and Andres Freund
Discussion: 2400449.GjM57CE0Yg@dinodell
This feature is controlled by a new old_snapshot_threshold GUC. A
value of -1 disables the feature, and that is the default. The
value of 0 is just intended for testing. Above that it is the
number of minutes a snapshot can reach before pruning and vacuum
are allowed to remove dead tuples which the snapshot would
otherwise protect. The xmin associated with a transaction ID does
still protect dead tuples. A connection which is using an "old"
snapshot does not get an error unless it accesses a page modified
recently enough that it might not be able to produce accurate
results.
This is similar to the Oracle feature, and we use the same SQLSTATE
and error message for compatibility.
This patch is a no-op patch which is intended to reduce the chances
of failures of omission once the functional part of the "snapshot
too old" patch goes in. It adds parameters for snapshot, relation,
and an enum to specify whether the snapshot age check needs to be
done for the page at this point. This initial patch passes NULL
for the first two new parameters and BGP_NO_SNAPSHOT_TEST for the
third. The follow-on patch will change the places where the test
needs to be made.
Contention on the relation extension lock can become quite fierce when
multiple processes are inserting data into the same relation at the same
time at a high rate. Experimentation shows the extending the relation
multiple blocks at a time improves scalability.
Dilip Kumar, reviewed by Petr Jelinek, Amit Kapila, and me.
Commit ac1d794 ("Make idle backends exit if the postmaster dies.")
introduced a regression on, at least, large linux systems. Constantly
adding the same postmaster_alive_fds to the OSs internal datastructures
for implementing poll/select can cause significant contention; leading
to a performance regression of nearly 3x in one example.
This can be avoided by using e.g. linux' epoll, which avoids having to
add/remove file descriptors to the wait datastructures at a high rate.
Unfortunately the current latch interface makes it hard to allocate any
persistent per-backend resources.
Replace, with a backward compatibility layer, WaitLatchOrSocket with a
new WaitEventSet API. Users can allocate such a Set across multiple
calls, and add more than one file-descriptor to wait on. The latter has
been added because there's upcoming postgres features where that will be
helpful.
In addition to the previously existing poll(2), select(2),
WaitForMultipleObjects() implementations also provide an epoll_wait(2)
based implementation to address the aforementioned performance
problem. Epoll is only available on linux, but that is the most likely
OS for machines large enough (four sockets) to reproduce the problem.
To actually address the aforementioned regression, create and use a
long-lived WaitEventSet for FE/BE communication. There are additional
places that would benefit from a long-lived set, but that's a task for
another day.
Thanks to Amit Kapila, who helped make the windows code I blindly wrote
actually work.
Reported-By: Dmitry Vasilyev Discussion:
CAB-SwXZh44_2ybvS5Z67p_CDz=XFn4hNAD=CnMEF+QqkXwFrGg@mail.gmail.com20160114143931.GG10941@awork2.anarazel.de
Previously latches for windows and unix had been implemented in
different files. A later patch introduce an expanded wait
infrastructure, keeping the implementation separate would introduce too
much duplication.
This basically just moves the functions, without too much change. The
reason to keep this separate is that it allows blame to continue working
a little less badly; and to make review a tiny bit easier.
Discussion: 20160114143931.GG10941@awork2.anarazel.de
Up to now checkpoints were written in the order they're in the
BufferDescriptors. That's nearly random in a lot of cases, which
performs badly on rotating media, but even on SSDs it causes slowdowns.
To avoid that, sort checkpoints before writing them out. We currently
sort by tablespace, relfilenode, fork and block number.
One of the major reasons that previously wasn't done, was fear of
imbalance between tablespaces. To address that balance writes between
tablespaces.
The other prime concern was that the relatively large allocation to sort
the buffers in might fail, preventing checkpoints from happening. Thus
pre-allocate the required memory in shared memory, at server startup.
This particularly makes it more efficient to have checkpoint flushing
enabled, because that'll often result in a lot of writes that can be
coalesced into one flush.
Discussion: alpine.DEB.2.10.1506011320000.28433@sto
Author: Fabien Coelho and Andres Freund
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; especially on filesystems like xfs and ext4 when mounted
with data=writeback. To be certain that a rename() atomically replaces
the previous file contents in the face of crashes and different
filesystems, one has to fsync the old filename, rename the file, fsync
the new filename, fsync the containing directory. This sequence is not
generally adhered to currently; which exposes us to data loss risks. To
avoid having to repeat this arduous sequence, introduce
durable_rename(), which wraps all that.
Also add durable_link_or_rename(). Several places use link() (with a
fallback to rename()) to rename a file, trying to avoid replacing the
target file out of paranoia. Some of those rename sequences need to be
durable as well. There seems little reason extend several copies of the
same logic, so centralize the link() callers.
This commit does not yet make use of the new functions; they're used in
a followup commit.
Author: Michael Paquier, Andres Freund
Discussion: 56583BDD.9060302@2ndquadrant.com
Backpatch: All supported branches
For example, if you want to perform an ioctl() on a file descriptor
opened through the fd.c routines, there's no way to do that without
being able to get at the underlying fd.
KaiGai Kohei
Commit a892234f83 added a second bit per
page to the visibility map, which still seems like a good idea, but it
also added a second page-level bit alongside PD_ALL_VISIBLE to track
whether the visibility map bit was set. That no longer seems like a
clever plan, because we don't really need that bit for anything. We
always clear both bits when the page is modified anyway.
Patch by me, reviewed by Kyotaro Horiguchi and Masahiko Sawada.
The new bit indicates whether every tuple on the page is already frozen.
It is cleared only when the all-visible bit is cleared, and it can be
set only when we vacuum a page and find that every tuple on that page is
both visible to every transaction and in no need of any future
vacuuming.
A future commit will use this new bit to optimize away full-table scans
that would otherwise be triggered by XID wraparound considerations. A
page which is merely all-visible must still be scanned in that case, but
a page which is all-frozen need not be. This commit does not attempt
that optimization, although that optimization is the goal here. It
seems better to get the basic infrastructure in place first.
Per discussion, it's very desirable for pg_upgrade to automatically
migrate existing VM forks from the old format to the new format. That,
too, will be handled in a follow-on patch.
Masahiko Sawada, reviewed by Kyotaro Horiguchi, Fujii Masao, Amit
Kapila, Simon Riggs, Andres Freund, and others, and substantially
revised by me.
This patch introduces "pg_blocking_pids(int) returns int[]", which returns
the PIDs of any sessions that are blocking the session with the given PID.
Historically people have obtained such information using a self-join on
the pg_locks view, but it's unreasonably tedious to do it that way with any
modicum of correctness, and the addition of parallel queries has pretty
much broken that approach altogether. (Given some more columns in the view
than there are today, you could imagine handling parallel-query cases with
a 4-way join; but ugh.)
The new function has the following behaviors that are painful or impossible
to get right via pg_locks:
1. Correctly understands which lock modes block which other ones.
2. In soft-block situations (two processes both waiting for conflicting lock
modes), only the one that's in front in the wait queue is reported to
block the other.
3. In parallel-query cases, reports all sessions blocking any member of
the given PID's lock group, and reports a session by naming its leader
process's PID, which will be the pg_backend_pid() value visible to
clients.
The motivation for doing this right now is mostly to fix the isolation
tests. Commit 38f8bdcac4 lobotomized
isolationtester's is-it-waiting query by removing its ability to recognize
nonconflicting lock modes, as a crude workaround for the inability to
handle soft-block situations properly. But even without the lock mode
tests, the old query was excessively slow, particularly in
CLOBBER_CACHE_ALWAYS builds; some of our buildfarm animals fail the new
deadlock-hard test because the deadlock timeout elapses before they can
probe the waiting status of all eight sessions. Replacing the pg_locks
self-join with use of pg_blocking_pids() is not only much more correct, but
a lot faster: I measure it at about 9X faster in a typical dev build with
Asserts, and 3X faster in CLOBBER_CACHE_ALWAYS builds. That should provide
enough headroom for the slower CLOBBER_CACHE_ALWAYS animals to pass the
test, without having to lengthen deadlock_timeout yet more and thus slow
down the test for everyone else.
We don't really need this field, because it's either zero or redundant with
PGPROC.pid. The use of zero to mark "not a group leader" is not necessary
since we can just as well test whether lockGroupLeader is NULL. This does
not save very much, either as to code or data, but the simplification seems
worthwhile anyway.
Reflow text in lock manager README so that it fits within 80 columns.
Correct some mistakes. Expand the README to explain not only why group
locking exists but also the data structures that support it. Improve
comments related to group locking several files. Change the name of a
macro argument for improved clarity.
Most of these problems were reported by Tom Lane, but I found a few
of them myself.
Robert Haas and Tom Lane
This finishes the work - spread across many commits over the last
several months - of putting each type of lock other than the named
individual locks into a separate tranche.
Amit Kapila
Commit 0e141c0fbb introduced a new
facility to reduce ProcArrayLock contention by clearing several XIDs
from the ProcArray under a single lock acquisition. The names
initially chosen were deemed not to be very good choices, so commit
4aec49899e renamed them. But now it
seems like we still didn't get it right. A pending patch wants to
add similar infrastructure for batching CLOG updates, so the names
need to be clear enough to allow a new set of structure members with
a related purpose.
Amit Kapila
As of commit c1772ad922, there's no
longer any way of requesting additional LWLocks in the main tranche,
so we don't need NumLWLocks() or LWLockAssign() any more. Also,
some of the allocation counters that we had previously aren't needed
any more either.
Amit Kapila
For locking purposes, we now regard heavyweight locks as mutually
non-conflicting between cooperating parallel processes. There are some
possible pitfalls to this approach that are not to be taken lightly,
but it works OK for now and can be changed later if we find a better
approach. Without this, it's very easy for parallel queries to
silently self-deadlock if the user backend holds strong relation locks.
Robert Haas, with help from Amit Kapila. Thanks to Noah Misch and
Andres Freund for extensive discussion of possible issues with this
approach.
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
This is following in a long train of similar changes and for the same
reasons - see b319356f0e and
fe702a7b3f inter alia.
Author: Amit Kapila
Reviewed-by: Alexander Korotkov, Robert Haas
Previously, each PGPROC's backendLock was part of the main tranche,
and the PGPROC just contained a pointer. Now, the actual LWLock is
part of the PGPROC.
As with previous, similar patches, this makes it significantly easier
to identify these lwlocks in LWLOCK_STATS or Trace_lwlocks output
and improves modularity.
Author: Ildus Kurbangaliev
Reviewed-by: Amit Kapila, Robert Haas
Move the content lock directly into the BufferDesc, so that locking and
pinning a buffer touches only one cache line rather than two. Adjust
the definition of BufferDesc slightly so that this doesn't make the
BufferDesc any larger than one cache line (at least on platforms where
a spinlock is only 1 or 2 bytes).
We can't fit the I/O locks into the BufferDesc and stay within one
cache line, so move those to a completely separate tranche. This
leaves a relatively limited number of LWLocks in the main tranche, so
increase the padding of those remaining locks to a full cache line,
rather than allowing adjacent locks to share a cache line, hopefully
reducing false sharing.
Performance testing shows that these changes make little difference
on laptop-class machines, but help significantly on larger servers,
especially those with more than 2 sockets.
Andres Freund, originally based on an earlier patch by Simon Riggs.
Review and cosmetic adjustments (including heavy rewriting of the
comments) by me.
It's a bit cumbersome to use LWLockNewTrancheId(), because the returned
value needs to be shared between backends so that each backend can call
LWLockRegisterTranche() with the correct ID. So, for built-in tranches,
use a hard-coded value instead.
This is motivated by an upcoming patch adding further built-in tranches.
Andres Freund and Robert Haas
At the end of crash recovery, unlogged relations are reset to the empty
state, using their init fork as the template. The init fork is copied to
the main fork without going through shared buffers. Unfortunately WAL
replay so far has not necessarily flushed writes from shared buffers to
disk at that point. In normal crash recovery, and before the
introduction of 'fast promotions' in fd4ced523 / 9.3, the
END_OF_RECOVERY checkpoint flushes the buffers out in time. But with
fast promotions that's not the case anymore.
To fix, force WAL writes targeting the init fork to be flushed
immediately (using the new FlushOneBuffer() function). In 9.5+ that
flush can centrally be triggered from the code dealing with restoring
full page writes (XLogReadBufferForRedoExtended), in earlier releases
that responsibility is in the hands of XLOG_HEAP_NEWPAGE's replay
function.
Backpatch to 9.1, even if this currently is only known to trigger in
9.3+. Flushing earlier is more robust, and it is advantageous to keep
the branches similar.
Typical symptoms of this bug are errors like
'ERROR: index "..." contains unexpected zero page at block 0'
shortly after promoting a node.
Reported-By: Thom Brown
Author: Andres Freund and Michael Paquier
Discussion: 20150326175024.GJ451@alap3.anarazel.de
Backpatch: 9.1-
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.
Review by Andres Freund
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.
The shm_mq mechanism was built to send error (and notice) messages and
tuples between backends. However, shm_mq itself only deals in raw
bytes. Since commit 2bd9e412f9, we have
had infrastructure for one message to redirect protocol messages to a
queue and for another backend to parse them and do useful things with
them. This commit introduces a somewhat analogous facility for tuples
by adding a new type of DestReceiver, DestTupleQueue, which writes
each tuple generated by a query into a shm_mq, and a new
TupleQueueFunnel facility which reads raw tuples out of the queue and
reconstructs the HeapTuple format expected by the executor.
The TupleQueueFunnel abstraction supports reading from multiple tuple
streams at the same time, but only in round-robin fashion. Someone
could imaginably want other policies, but this should be good enough
to meet our short-term needs related to parallel query, and we can
always extend it later.
This also makes one minor addition to the shm_mq API that didn'
seem worth breaking out as a separate patch.
Extracted from Amit Kapila's parallel sequential scan patch. This
code was originally written by me, and then it was revised by Amit,
and then it was revised some more by me.
Naming the individual lwlocks seems like something that may be useful
for other types of debugging, monitoring, or instrumentation output,
but this commit just implements it for the specific case of
trace_lwlocks.
Patch by me, reviewed by Amit Kapila and Kyotaro Horiguchi
Per discussion, nowadays it is possible to have tablespaces that have
wildly different I/O characteristics from others. Setting different
effective_io_concurrency parameters for those has been measured to
improve performance.
Author: Julien Rouhaud
Reviewed by: Andres Freund
Post-commit review by Andres Freund discovered a couple of concurrency
bugs in the original patch: specifically, if the leader cleared a
follower's XID before it reached PGSemaphoreLock, the semaphore would be
left in the wrong state; and if another process did PGSemaphoreUnlock
for some unrelated reason, we might resume execution before the fact
that our XID was cleared was globally visible.
Also, improve the wording of some comments, rename nextClearXidElem
to firstClearXidElem in PROC_HDR for clarity, and drop some volatile
qualifiers that aren't necessary.
Amit Kapila, reviewed and slightly revised by me.
Some googling turned up multiple sources saying that older versions of icc
do not accept gcc-compatible asm blocks on IA64, though asm does work on
x86[_64]. This is apparently fixed as of icc version 12.0 or so, but that
doesn't help us much; if we have to carry the extra implementation anyway,
we may as well just use it for icc rather than add a compiler version test.
Hence, revert commit 2c713d6ea2 (though I
separated the icc code from the gcc code completely, producing what seems
cleaner code). Document the state of affairs more explicitly, both in
s_lock.h and postgres.c, and make some cosmetic adjustments around the
IA64 code in s_lock.h.
On recent AIX it's necessary to configure gcc to use the native assembler
(because the GNU assembler hasn't been updated to handle AIX 6+). This
caused PG builds to fail with assembler syntax errors, because we'd try
to compile s_lock.h's gcc asm fragment for PPC, and that assembly code
relied on GNU-style local labels. We can't substitute normal labels
because it would fail in any file containing more than one inlined use of
tas(). Fortunately, that code is stable enough, and the PPC ISA is simple
enough, that it doesn't seem like too much of a maintenance burden to just
hand-code the branch offsets, removing the need for any labels.
Note that the AIX assembler only accepts "$" for the location counter
pseudo-symbol. The usual GNU convention is "."; but it appears that all
versions of gas for PPC also accept "$", so in theory this patch will not
break any other PPC platforms.
This has been reported by a few people, but Steve Underwood gets the credit
for being the first to pursue the problem far enough to understand why it
was failing. Thanks also to Noah Misch for additional testing.
Some frontend code like e.g. pg_xlogdump or pg_resetxlog, has to use
backend headers. Unfortunately until now that code includes most of the
locking code. It's generally not nice to expose such low level details,
but de6fd1c898 made that a hard problem. We fall back to defining
'inline' away if the compiler doesn't support it - that can cause linker
errors like on buildfarm animal pademelon if a inline function
references backend only code.
To fix that problem separate definitions from lock.h that are required
from frontend code into lockdefs.h and use it in the relevant
places. I've only removed the minimal amount of necessary definitions
for now - it might turn out that we want more for other reasons.
To avoid such details being exposed again put some checks against being
included from frontend code into atomics.h, lock.h, lwlock.h and
s_lock.h. It's otherwise fairly easy to indirectly include these
headers.
Discussion: 20150806070902.GE12214@awork2.anarazel.de
When a write transaction commits, it must clear its XID advertised via
the ProcArray, which requires that we hold ProcArrayLock in exclusive
mode in order to prevent concurrent processes running GetSnapshotData
from seeing inconsistent results. When many processes try to commit
at once, ProcArrayLock must change hands repeatedly, with each
concurrent process trying to commit waking up to acquire the lock in
turn. To make things more efficient, when more than one backend is
trying to commit a write transaction at the same time, have just one
of them acquire ProcArrayLock in exclusive mode and clear the XIDs of
all processes in the group. Benchmarking reveals that this is much
more efficient at very high client counts.
Amit Kapila, heavily revised by me, with some review also from Pavan
Deolasee.
The lwlock scalability work introduced two race conditions into the
lwlock variable support provided for xlog.c. First, and harmlessly on
most platforms, it set/read the variable without the spinlock in some
places. Secondly, due to the removal of the spinlock, it was possible
that a backend missed changes to the variable's state if it changed in
the wrong moment because checking the lock's state, the variable's state
and the queuing are not protected by a single spinlock acquisition
anymore.
To fix first move resetting the variable's from LWLockAcquireWithVar to
WALInsertLockRelease, via a new function LWLockReleaseClearVar. That
prevents issues around waiting for a variable's value to change when a
new locker has acquired the lock, but not yet set the value. Secondly
re-check that the variable hasn't changed after enqueing, that prevents
the issue that the lock has been released and already re-acquired by the
time the woken up backend checks for the lock's state.
Reported-By: Jeff Janes
Analyzed-By: Heikki Linnakangas
Reviewed-By: Heikki Linnakangas
Discussion: 5592DB35.2060401@iki.fi
Backpatch: 9.5, where the lwlock scalability went in
This code was originally written as part of parallel query effort, but
it seems to have independent value, because if we make one decision
about where to get a PGPROC when we allocate and then put it back on a
different list at backend-exit time, bad things happen. This isn't
just a theoretical risk; we fixed an actual problem of this type in
commit e280c630a8.
_Asm_sched_fence() is just a compiler barrier, not a memory barrier. But
spinlock release on IA64 needs, at the very least, release
semantics. Use a full barrier instead.
This might be the cause for the occasional failures on buildfarm member
anole.
Discussion: 20150629101108.GB17640@alap3.anarazel.de
Commit 2ce439f337 introduced a rather serious
regression, namely that if its scan of the data directory came across any
un-fsync-able files, it would fail and thereby prevent database startup.
Worse yet, symlinks to such files also caused the problem, which meant that
crash restart was guaranteed to fail on certain common installations such
as older Debian.
After discussion, we agreed that (1) failure to start is worse than any
consequence of not fsync'ing is likely to be, therefore treat all errors
in this code as nonfatal; (2) we should not chase symlinks other than
those that are expected to exist, namely pg_xlog/ and tablespace links
under pg_tblspc/. The latter restriction avoids possibly fsync'ing a
much larger part of the filesystem than intended, if the user has left
random symlinks hanging about in the data directory.
This commit takes care of that and also does some code beautification,
mainly moving the relevant code into fd.c, which seems a much better place
for it than xlog.c, and making sure that the conditional compilation for
the pre_sync_fname pass has something to do with whether pg_flush_data
works.
I also relocated the call site in xlog.c down a few lines; it seems a
bit silly to be doing this before ValidateXLOGDirectoryStructure().
The similar logic in initdb.c ought to be made to match this, but that
change is noncritical and will be dealt with separately.
Back-patch to all active branches, like the prior commit.
Abhijit Menon-Sen and Tom Lane
Robert Haas
Tom Lane
Andres Freund
Simon Riggs
Kevin Grittner
Greg Stark
Alvaro Herrera
Magnus Hagander
Joe Conway
Bruce Momjian
Heikki Linnakangas