Instead of letting every backend participating in a group commit wait
independently, have the first one that becomes ready to flush WAL wait
for the configured delay, and let all the others wait just long enough
for that first process to complete its flush. This greatly increases
the chances of being able to configure a commit_delay setting that
actually improves performance.
As a side consequence of this change, commit_delay now affects all WAL
flushes, rather than just commits. There was some discussion on
pgsql-hackers about whether to rename the GUC to, say, wal_flush_delay,
but in the absence of consensus I am leaving it alone for now.
Peter Geoghegan, with some changes, mostly to the documentation, by me.
Per testing by Andres Freund, this improves replication performance
and reduces replication latency and latency jitter. I was a bit
concerned about moving more work into XLogInsert, but testing seems
to show that it's not a problem in practice.
Along the way, improve comments for WaitLatchOrSocket.
Andres Freund. Review and stylistic cleanup by me.
If the record header is garbled, we're now quite likely to notice it before
we try to make a bogus memory allocation and run out of memory. That can
still happen, if the xlog record is split across pages (we cannot verify
the record header until reading the next page in that scenario), but this
reduces the chances. An out-of-memory is treated as a corrupt record
anyway, so this isn't a correctness issue, just a case of giving a better
error message.
Per Amit Kapila's suggestion.
The original coding failed miserably for BLCKSZ of 4K or less, as reported
by Josh Kupershmidt. With the present design for text indexes, a given
inner tuple could have up to 256 labels (requiring either 3K or 4K bytes
depending on MAXALIGN), which means that we can't positively guarantee no
failures for smaller blocksizes. But we can at least make it behave sanely
so long as there are few enough labels to fit on a page. Considering that
btree is also more prone to "index tuple too large" failures when BLCKSZ is
small, it's not clear that we should expend more work than this on this
case.
Avoid using LockPage(rel, 0, lockmode) to protect against changes to
the bucket mapping. Instead, an exclusive buffer content lock is now
viewed as sufficient permission to modify the metapage, and a shared
buffer content lock is used when such modifications need to be
prevented. This more relaxed locking regimen makes it possible that,
when we're busy getting a heavyweight bucket on the bucket we intend
to search or insert into, a bucket split might occur underneath us.
To compenate for that possibility, we use a loop-and-retry system:
release the metapage content lock, acquire the heavyweight lock on the
target bucket, and then reacquire the metapage content lock and check
that the bucket mapping has not changed. Normally it hasn't, and
we're done. But if by chance it has, we simply unlock the metapage,
release the heavyweight lock we acquired previously, lock the new
bucket, and loop around again. Even in the worst case we cannot loop
very many times here, since we don't split the same bucket again until
we've split all the other buckets, and 2^N gets big pretty fast.
This results in greatly improved concurrency, because we're
effectively replacing two lwlock acquire-and-release cycles in
exclusive mode (on one of the lock manager locks) with a single
acquire-and-release cycle in shared mode (on the metapage buffer
content lock). Testing shows that it's still not quite as good as
btree; for that, we'd probably have to find some way of getting rid
of the heavyweight bucket locks as well, which does not appear
straightforward.
Patch by me, review by Jeff Janes.
The latter was already the dominant use, and it's preferable because
in C the convention is that intXX means XX bits. Therefore, allowing
mixed use of int2, int4, int8, int16, int32 is obviously confusing.
Remove the typedefs for int2 and int4 for now. They don't seem to be
widely used outside of the PostgreSQL source tree, and the few uses
can probably be cleaned up by the time this ships.
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 saves a few bytes of WAL space, but the real motivation is to make it
predictable how much WAL space a record requires, as it no longer depends
on whether we need to waste the last few bytes at end of WAL page because
the header doesn't fit.
The total length field of WAL record, xl_tot_len, is moved to the beginning
of the WAL record header, so that it is still always found on the first page
where a WAL record begins.
Bump WAL version number again as this is an incompatible change.
The continuation record only contained one field, xl_rem_len, so it makes
things simpler to just include it in the WAL page header. This wastes four
bytes on pages that don't begin with a continuation from previos page, plus
four bytes on every page, because of padding.
The motivation of this is to make it easier to calculate how much space a
WAL record needs. Before this patch, it depended on how many page boundaries
the record crosses. The motivation of that, in turn, is to separate the
allocation of space in the WAL from the copying of the record data to the
allocated space. Keeping the calculation of space required simple helps to
keep the critical section of allocating the space from WAL short. But that's
not included in this patch yet.
Bump WAL version number again, as this is an incompatible change.
The comments claimed that wasting the last segment made it easier to do
calculations with XLogRecPtrs, because you don't have problems representing
last-byte-position-plus-1 that way. In my experience, however, it only made
things more complicated, because the there was two ways to represent the
boundary at the beginning of a logical log file: logid = n+1 and xrecoff = 0,
or as xlogid = n and xrecoff = 4GB - XLOG_SEG_SIZE. Some functions were
picky about which representation was used.
Also, use a 64-bit segment number instead of the log/seg combination, to
point to a certain WAL segment. We assume that all platforms have a working
64-bit integer type nowadays.
This is an incompatible change in WAL format, so bumping WAL version number.
Because permissions are assigned to element types, not array types,
complaining about permission denied on an array type would be
misleading to users. So adjust the reporting to refer to the element
type instead.
In order not to duplicate the required logic in two dozen places,
refactor the permission denied reporting for types a bit.
pointed out by Yeb Havinga during the review of the type privilege
feature
There's no need to hold onto the locks until the files are needed,
and by doing it this way, we reduce the impact on other backends who
may be awaiting locks we hold.
Noah Misch
When I implemented the ginbuildempty() function as part of
implementing unlogged tables, I falsified the note in the header
comment for log_newpage. Although we could fix that up by changing
the comment, it seems cleaner to add a new function which is
specifically intended to handle this case. So do that.
This reverts commit 18fb9d8d21. Per
discussion, it does not seem like a good idea to allow committed changes to
go un-checkpointed indefinitely, as could happen in a low-traffic server;
that makes us entirely reliant on the WAL stream with no redundancy that
might aid data recovery in case of disk failure.
This re-introduces the original problem of hot-standby setups generating a
small continuing stream of WAL traffic even when idle, but there are other
ways to address that without compromising crash recovery, so we'll revisit
that issue in a future release cycle.
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
WALSender now woken up after each background flush by WALwriter, avoiding
multi-second replication delay for an all-async commit workload.
Replication delay reduced from 7s with default settings to 200ms and often
much less, allowing significantly reduced data loss at failover.
Andres Freund and Simon Riggs
In lazy_scan_heap, we could issue bogus warnings about incorrect
information in the visibility map, because we checked the visibility
map bit before locking the heap page, creating a race condition. Fix
by rechecking the visibility map bit before we complain. Rejigger
some related logic so that we rely on the possibly-outdated
all_visible_according_to_vm value as little as possible.
In heap_multi_insert, it's not safe to clear the visibility map bit
before beginning the critical section. The visibility map is not
crash-safe unless we treat clearing the bit as a critical operation.
Specifically, if the transaction were to error out after we set the
bit and before entering the critical section, we could end up writing
the heap page to disk (with the bit cleared) and crashing before the
visibility map page made it to disk. That would be bad. heap_insert
has this correct, but somehow the order of operations got rearranged
when heap_multi_insert was added.
Also, add some more comments to visibilitymap_test, lazy_scan_heap,
and IndexOnlyNext, expounding on concurrency issues.
Per extensive code review by Andres Freund, and further review by Tom
Lane, who also made the original report about the bogus warnings.
When we allowed read-only transactions to skip assigning XIDs
we introduced the possibility that a fully deleted btree page
could be reused. This broke the index link sequence which could
then lead to indexscans silently returning fewer rows than would
have been correct. The actual incidence of silent errors from
this is thought to be very low because of the exact workload
required and locking pre-conditions. Fix is to remove pages only
if index page opaque->btpo.xact precedes RecentGlobalXmin.
Noah Misch, reviewed by Simon Riggs
We used to mimic the way a stack is constructed when descending the tree
during normal GiST inserts, but that was quite complicated during a buffered
build. It was also wrong: in GiST, the left-to-right relationships on
different levels might not match each other, so that when you know the
parent of a child page, you won't necessarily find the parent of the page to
the right of the child page by following the rightlinks at the parent level.
This sometimes led to "could not re-find parent" errors while building a
GiST index.
We now use a simple hash table to track the parent of every internal page.
Whenever a page is split, and downlinks are moved from one page to another,
we update the hash table accordingly. This is also better for performance
than the old method, as we never need to move right to re-find the parent
page, which could take a significant amount of time for buffers that were
created much earlier in the index build.
There were two bugs here: We forgot to call gistFreeBuildBuffers() function
at the end of build, and we passed interXact == true to BufFileCreateTemp,
so the file wasn't automatically cleaned up at end-of-transaction either.
The result of (maintenance_work_mem * 1024) / BLCKSZ doesn't fit in a signed
32-bit integer, if maintenance_work_mem >= 2GB. Use double instead. And
while we're at it, write the calculations in an easier to understand form,
with the intermediary steps written out and commented.
AbortOutOfAnyTransaction failed to do anything if the state it saw on
entry corresponded to failing partway through StartTransaction. I fixed
AbortCurrentTransaction to cope with that case way back in commit
60b2444cc3, but evidently overlooked that
AbortOutOfAnyTransaction should do likewise.
Back-patch to all supported branches. It's not clear that this omission
has any more-than-cosmetic consequences, but it's also not clear that it
doesn't, so back-patching seems the least risky choice.
The only interesting-for-performance case wherein we force heapscan here
is when we're rebuilding the relcache init file, and the only such case
that is likely to be examining a catalog big enough to be syncscanned is
RelationBuildTupleDesc. But the early-exit optimization in that code gets
broken if we start the scan at a random place within the catalog, so that
allowing syncscan is actually a big deoptimization if pg_attribute is large
(at least for the normal case where the rows for core system catalogs have
never been changed since initdb). Hence, prevent syncscan here. Per my
testing pursuant to complaints from Jeff Frost and Greg Sabino Mullane,
though neither of them seem to have actually hit this specific problem.
Back-patch to 8.3, where syncscan was introduced.
If a seqscan encounters many consecutive pages containing only dead tuples,
it can remain in the loop in heapgettup for a long time, and there was no
CHECK_FOR_INTERRUPTS anywhere in that loop. This meant there were
real-world situations where a query would be effectively uncancelable for
long stretches. Add a check placed to occur once per page, which should be
enough to provide reasonable response time without adding any measurable
overhead.
Report and patch by Merlin Moncure (though I tweaked it a bit).
Back-patch to all supported branches.
When we create a temporary copy of the old node buffer, in stack, we mustn't
leak that into any of the long-lived data structures. Before this patch,
when we called gistPopItupFromNodeBuffer(), it got added to the array of
"loaded buffers". After gistRelocateBuildBuffersOnSplit() exits, the
pointer added to the loaded buffers array points to garbage. Often that goes
unnotied, because when we go through the array of loaded buffers to unload
them, buffers with a NULL pageBuffer are ignored, which can often happen by
accident even if the pointer points to garbage.
This patch fixes that by marking the temporary copy in stack explicitly as
temporary, and refrain from adding buffers marked as temporary to the array
of loaded buffers.
While we're at it, initialize nodeBuffer->pageBlocknum to InvalidBlockNumber
and improve comments a bit. This isn't strictly necessary, but makes
debugging easier.
If the amount of freespace on page was less than the amount reserved by
fillfactor, the calculation would underflow.
This fixes bug #6643 reported by Tomonari Katsumata.
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
When inserting the downlinks for a split gist page, we used hold the locks
on the child pages until the insertion into the parent - and recursively its
parent if it had to be split too - were all completed. Change that so that
the locks on child pages are released after the insertion in the immediate
parent is done, before recursing further up the tree.
This reduces the number of lwlocks that are held simultaneously. Holding
many locks is bad for concurrency, and in extreme cases you can even hit
the limit of 100 simultaneously held lwlocks in a backend. If you're really
unlucky, you can hit the limit while in a critical section, which brings
down the whole system.
This fixes bug #6629 reported by Tom Forbes. Backpatch to 9.1. The page
splitting code was rewritten in 9.1, and the old code did not have this
problem.
Users of asynchronous-commit mode expect there to be a guaranteed maximum
delay before an async commit's WAL records get flushed to disk. The
original version of the walwriter hibernation patch broke that. Add an
extra shared-memory flag to allow async commits to kick the walwriter out
of hibernation mode, without adding any noticeable overhead in cases where
no action is needed.
This patch modifies the walwriter process so that, when it has not found
anything useful to do for many consecutive wakeup cycles, it extends its
sleep time to reduce the server's idle power consumption. It reverts to
normal as soon as it's done any successful flushes. It's still true that
during any async commit, backends check for completed, unflushed pages of
WAL and signal the walwriter if there are any; so that in practice the
walwriter can get awakened and returned to normal operation sooner than the
sleep time might suggest.
Also, improve the checkpointer so that it uses a latch and a computed delay
time to not wake up at all except when it has something to do, replacing a
previous hardcoded 0.5 sec wakeup cycle. This also is primarily useful for
reducing the server's power consumption when idle.
In passing, get rid of the dedicated latch for signaling the walwriter in
favor of using its procLatch, since that comports better with possible
generic signal handlers using that latch. Also, fix a pre-existing bug
with failure to save/restore errno in walwriter's signal handlers.
Peter Geoghegan, somewhat simplified by Tom
At the time we check whether the tuple is dead to all running
transactions, we've already verified that it isn't visible to our
scan, setting hint bits if appropriate. So there's no need to
recheck CLOG for the all-dead test we do just a moment later.
So, add HeapTupleIsSurelyDead() to test the appropriate condition
under the assumption that all relevant hit bits are already set.
Review by Tom Lane.
Peter Eisentraut
Robert Haas
Heikki Linnakangas
Tom Lane
Alvaro Herrera
Bruce Momjian
Simon Riggs