Such paths are unsafe. To make it cheaper to detect when this case
applies, track whether a relation's default PathTarget contains any
non-Vars. In most cases, the answer will be no, which enables us to
determine cheaply that the target list for a proposed path is
parallel-safe. However, subquery pull-up can create cases that
require us to inspect the target list more carefully.
Amit Kapila, reviewed by me.
Document these as "nearest integer >= argument" and "nearest integer <=
argument", which will hopefully be less confusing than the old formulation.
New wording is from Matlab via Dean Rasheed.
I changed the pg_description entries as well as the SGML docs. In the
back branches, this will only affect installations initdb'd in the future,
but it should be harmless otherwise.
Discussion: <CAEZATCW3yzJo-NMSiQs5jXNFbTsCEftZS-Og8=FvFdiU+kYuSA@mail.gmail.com>
Fix a couple of overlooked uses of "degree" terminology. Make the parallel
worker count selection logic in create_plain_partial_paths more robust (in
particular, it failed with max_parallel_workers_per_gather set to zero).
This terminology provoked widespread complaints. So, instead, rename
the GUC max_parallel_degree to max_parallel_workers_per_gather
(leaving room for a possible future GUC max_parallel_workers that acts
as a system-wide limit), and rename the parallel_degree reloption to
parallel_workers. Rename structure members to match.
These changes create a dump/restore hazard for users of PostgreSQL
9.6beta1 who have set the reloption (or applied the GUC using ALTER
USER or ALTER DATABASE).
This commit reverts 137805f89 as well as the associated commits 015e88942,
5306df283, and 68d704edb. We found multiple bugs in this feature, and
there was concern about possible planner slowdown (though to be fair,
exhibiting a very large slowdown proved difficult). The way forward
requires a considerable rewrite, which may or may not be possible to
accomplish in time for beta2. In my judgment reviewing the rewrite will
be easier to accomplish starting from a clean slate, so let's temporarily
revert what's there now. This also leaves us in a safe state if it turns
out to be necessary to postpone the rewrite to the next development cycle.
Discussion: <20160429102531.GA13701@huehner.biz>
If we ANALYZE only selected columns of a table, we should not postpone
auto-analyze because of that; other columns may well still need stats
updates. As committed, the counter is left alone if a column list is
given, whether or not it includes all analyzable columns of the table.
Per complaint from Tomasz Ostrowski.
It's been like this a long time, so back-patch to all supported branches.
Report: <ef99c1bd-ff60-5f32-2733-c7b504eb960c@ato.waw.pl>
If a Gather node has read as many tuples as it needs (for example, due
to Limit) it may detach the queue connecting it to the worker before
reading all of the worker's tuples. Rather than let the worker
continue to generate and send all of the results, have it stop after
sending the next tuple.
More could be done here to stop the worker even quicker, but this is
about as well as we can hope to do for 9.6.
This is in response to a problem report from Andreas Seltenreich.
Commit 44339b892a should be actually be
sufficient to fix that example even without this change, but it seems
better to do this, too, since we might otherwise waste quite a large
amount of effort in one or more workers.
Discussion: CAA4eK1KOKGqmz9bGu+Z42qhRwMbm4R5rfnqsLCNqFs9j14jzEA@mail.gmail.com
Amit Kapila
This attempts to buy back some of whatever performance we lost from fixing
bug #14174 by inlining the initial checks in MakeExpandedObjectReadOnly()
into the callers. We can do that in a macro without creating multiple-
evaluation hazards, so it's pretty much free notationally; and the amount
of code added to callers should be minimal as well. (Testing a value can't
take many more instructions than passing it to a subroutine.)
Might as well inline DatumIsReadWriteExpandedObject() while we're at it.
This is an ABI break for callers, so it doesn't seem safe to put into 9.5,
but I see no reason not to do it in HEAD.
If a plan node output expression returns an "expanded" datum, and that
output column is referenced in more than one place in upper-level plan
nodes, we need to ensure that what is returned is a read-only reference
not a read/write reference. Otherwise one of the referencing sites could
scribble on or even delete the expanded datum before we have evaluated the
others. Commit 1dc5ebc907, which introduced this feature, supposed
that it'd be sufficient to make SubqueryScan nodes force their output
columns to read-only state. The folly of that was revealed by bug #14174
from Andrew Gierth, and really should have been immediately obvious
considering that the planner will happily optimize SubqueryScan nodes
out of the plan without any regard for this issue.
The safest fix seems to be to make ExecProject() force its results into
read-only state; that will cover every case where a plan node returns
expression results. Actually we can delegate this to ExecTargetList()
since we can recursively assume that plain Vars will not reference
read-write datums. That should keep the extra overhead down to something
minimal. We no longer need ExecMakeSlotContentsReadOnly(), which was
introduced only in support of the idea that just a few plan node types
would need to do this.
In the future it would be nice to have the planner account for this problem
and inject force-to-read-only expression evaluation nodes into only the
places where there's a risk of multiple evaluation. That's not a suitable
solution for 9.5 or even 9.6 at this point, though.
Report: <20160603124628.9932.41279@wrigleys.postgresql.org>
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
Per post-commit review comments from Andres Freund, improve variable
names, comments, and in one place, slightly improve the code structure.
Masahiko Sawada
The original intent in the stats collector was that we should not write out
stats data oftener than every PGSTAT_STAT_INTERVAL msec. Backends will not
make requests at all if they see the existing data is newer than that, and
the stats collector is supposed to disregard requests having a cutoff_time
older than its most recently written data, so that close-together requests
don't result in multiple writes. But the latter part of that got broken
in commit 187492b6c2, so that if two backends concurrently decide
the existing stats are too old, the collector would write the data twice.
(In principle the collector's logic would still merge requests as long as
the second one arrives before we've actually written data ... but since
the message collection loop would write data immediately after processing
a single inquiry message, that never happened in practice, and in any case
the window in which it might work would be much shorter than
PGSTAT_STAT_INTERVAL.)
To fix, improve pgstat_recv_inquiry so that it checks whether the cutoff
time is too old, and doesn't add a request to the queue if so. This means
that we do not need DBWriteRequest.request_time, because the decision is
taken before making a queue entry. And that means that we don't really
need the DBWriteRequest data structure at all; an OID list of database
OIDs will serve and allow removal of some rather verbose and crufty code.
In passing, improve the comments in this area, which have been rather
neglected. Also change backend_read_statsfile so that it's not silently
relying on MyDatabaseId to have some particular value in the autovacuum
launcher process. It accidentally worked as desired because MyDatabaseId
is zero in that process; but that does not seem like a dependency we want,
especially with no documentation about it.
Although this patch is mine, it turns out I'd rediscovered a known bug,
for which Tomas Vondra had already submitted a patch that's functionally
equivalent to the non-cosmetic aspects of this patch. Thanks to Tomas
for reviewing this version.
Back-patch to 9.3 where the bug was introduced.
Prior-Discussion: <1718942738eb65c8407fcd864883f4c8@fuzzy.cz>
Patch: <4625.1464202586@sss.pgh.pa.us>
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
If both timeout indicators are set when we arrive at ProcessInterrupts,
we've historically just reported "lock timeout". However, some buildfarm
members have been observed to fail isolationtester's timeouts test by
reporting "lock timeout" when the statement timeout was expected to fire
first. The cause seems to be that the process is allowed to sleep longer
than expected (probably due to heavy machine load) so that the lock
timeout happens before we reach the point of reporting the error, and
then this arbitrary tiebreak rule does the wrong thing. We can improve
matters by comparing the scheduled timeout times to decide which error
to report.
I had originally proposed greatly reducing the 1-second window between
the two timeouts in the test cases. On reflection that is a bad idea,
at least for the case where the lock timeout is expected to fire first,
because that would assume that it takes negligible time to get from
statement start to the beginning of the lock wait. Thus, this patch
doesn't completely remove the risk of test failures on slow machines.
Empirically, however, the case this handles is the one we are seeing
in the buildfarm. The explanation may be that the other case requires
the scheduler to take the CPU away from a busy process, whereas the
case fixed here only requires the scheduler to not give the CPU back
right away to a process that has been woken from a multi-second sleep
(and, perhaps, has been swapped out meanwhile).
Back-patch to 9.3 where the isolationtester timeouts test was added.
Discussion: <8693.1464314819@sss.pgh.pa.us>
If RAW_EXPRESSION_COVERAGE_TEST is defined, do a no-op tree walk over
every basic DML statement submitted to parse analysis. If we'd had this
in place earlier, bug #14153 would have been caught by buildfarm testing.
The difficulty is that raw_expression_tree_walker() is only used in
limited cases involving CTEs (particularly recursive ones), so it's
very easy for an oversight in it to not be noticed during testing of a
seemingly-unrelated feature.
The type of error we can expect to catch with this is complete omission
of a node type from raw_expression_tree_walker(), and perhaps also
recursion into a field that doesn't contain a node tree, though that
would be an unlikely mistake. It won't catch failure to add new fields
that need to be recursed into, unfortunately.
I'll go enable this on one or two of my own buildfarm animals once
bug #14153 is dealt with.
Discussion: <27861.1464040417@sss.pgh.pa.us>
do_text_output_multiline() would fail (typically with a null pointer
dereference crash) if its input string did not end with a newline. Such
cases do not arise in our current sources; but it certainly could happen
in future, or in extension code's usage of the function, so we should fix
it. To fix, replace "eol += len" with "eol = text + len".
While at it, make two cosmetic improvements: mark the input string const,
and rename the argument from "text" to "txt" to dodge pgindent strangeness
(since "text" is a typedef name).
Even though this problem is only latent at present, it seems like a good
idea to back-patch the fix, since it's a very simple/safe patch and it's
not out of the realm of possibility that we might in future back-patch
something that expects sane behavior from do_text_output_multiline().
Per report from Hao Lee.
Report: <CAGoxFiFPAGyPAJLcFxTB5cGhTW2yOVBDYeqDugYwV4dEd1L_Ag@mail.gmail.com>
This was overlooked in commit 473b93287, which introduced DROP ACCESS
METHOD. Although that command is restricted to superusers, we don't want
even superusers dropping the built-in methods; "DROP ACCESS METHOD btree"
in particular is unrecoverable from. Pin these objects in the same way
that other initdb-created objects are pinned.
I chose to bump catversion for this fix. That's not absolutely necessary
perhaps, but it will ensure that no 9.6 production systems are missing
the pin entries.
This patch essentially reverts commit 4c6780fd17, in favor of a much
simpler solution for the case where the new cluster would choose to create
a TOAST table but the old cluster doesn't have one: just don't create a
TOAST table.
The existing code failed in at least two different ways if the situation
arose: (1) ALTER TABLE RESET didn't grab an exclusive lock, so that the
lock sanity check in create_toast_table failed; (2) pg_upgrade did not
provide a pg_type OID for the new toast table, so that the crosscheck in
TypeCreate failed. While both these problems were introduced by later
patches, they show that the hack being used to cause TOAST table creation
is overwhelmingly fragile (and untested). I also note that before the
TypeCreate crosscheck was added, the code would have resulted in assigning
an indeterminate pg_type OID to the toast table, possibly causing a later
OID conflict in that catalog; so that it didn't really work even when
committed.
If we simply don't create a TOAST table, there will only be a problem if
the code tries to store a tuple that's wider than a page, and field
compression isn't sufficient to get it under a page. Given that the TOAST
creation threshold is intended to be about a quarter of a page, it's very
hard to believe that cross-version differences in the do-we-need-a-toast-
table heuristic could result in an observable problem. So let's just
follow the old version's conclusion about whether a TOAST table is needed.
(If we ever do change needs_toast_table() so much that this conclusion
doesn't apply, we can devise a solution at that time, and hopefully do
it in a less klugy way than 4c6780fd17 did.)
Back-patch to 9.3, like the previous patch.
Discussion: <8110.1462291671@sss.pgh.pa.us>
Hash indexes are not WAL-logged, and so do not maintain the LSN of
index pages. Since the "snapshot too old" feature counts on
detecting error conditions using the LSN of a table and all indexes
on it, this makes it impossible to safely do early vacuuming on any
table with a hash index, so add this to the tests for whether the
xid used to vacuum a table can be adjusted based on
old_snapshot_threshold.
While at it, add a paragraph to the docs for old_snapshot_threshold
which specifically mentions this and other aspects of the feature
which may otherwise surprise users.
Problem reported and patch reviewed by Amit Kapila
Move fmtReloptionsArray() from pg_dump.c to string_utils.c so that it
is available to other frontend code. In particular psql's \ev and \sv
commands need it to handle view reloptions. Also rename the function
to appendReloptionsArray(), which is a more accurate description of
what it does.
Author: Dean Rasheed
Reviewed-by: Peter Eisentraut
Discussion: http://www.postgresql.org/message-id/CAEZATCWZjCgKRyM-agE0p8ax15j9uyQoF=qew7D2xB6cF76T8A@mail.gmail.com
The similarity of the original names to SQL keywords seems like a bad
idea. Rename them before we're stuck with 'em forever.
In passing, minor code and docs cleanup.
Discussion: <4875.1462210058@sss.pgh.pa.us>
This reverts commits f07d18b6e9, 82c83b3372, 3a3b309041, and
24c5f1a103.
This feature has shown enough immaturity that it was deemed better to
rip it out before rushing some more fixes at the last minute. There are
discussions on larger changes in this area for the next release.
Conversion functions were previously marked as parallel-unsafe, since
that is the default, but in fact they are safe. Parallel-safe
functions defined in pg_proc.h and redefined in system_views.sql were
ending up as parallel-unsafe because the redeclarations were not
marked PARALLEL SAFE. While editing system_views.sql, mark ts_debug()
parallel safe also.
Andreas Karlsson
Back in 3b02ea4f07 I added some comments in various places to explain
how logical decoding and other things worked. Not all of the changes
were welcome, because they were misleading or wrong. This changes them
a little bit to make them more accurate.
Some other comments are also changed to be more accurate. Also, fix a
bunch of typos.
Author: Álvaro Herrera, Craig Ringer
Andres Freund reviewed some parts of this.
Commit 7117685461 made pg_start_backup
parallel-restricted rather than parallel-safe, because it now relies
on backend-private state that won't be synchronized with the parallel
worker. However, it didn't update pg_proc.h. Separately, Andreas
Karlsson observed that system_views.sql neglected to reiterate the
parallel-safety markings whe redefining various functions, including
this one; so add a PARALLEL RESTRICTED declaration there to match
the new value in pg_proc.h.
Given a three-or-more-way equivalence class, such as X.Y = Y.Y = Z.Z,
it was possible for the planner to omit one of the quals needed to
enforce that all members of the equivalence class are actually equal.
This only happened in the case of a parameterized join node for two
of the relations, that is a plan tree like
Nested Loop
-> Scan X
-> Nested Loop
-> Scan Y
-> Scan Z
Filter: Z.Z = X.X
The eclass machinery normally expects to apply X.X = Y.Y when those
two relations are joined, but in this shape of plan tree they aren't
joined until the top node --- and, if the lower nested loop is marked
as parameterized by X, the top node will assume that the relevant eclass
condition(s) got pushed down into the lower node. On the other hand,
the scan of Z assumes that it's only responsible for constraining Z.Z
to match any one of the other eclass members. So one or another of
the required quals sometimes fell between the cracks, depending on
whether consideration of the eclass in get_joinrel_parampathinfo()
for the lower nested loop chanced to generate X.X = Y.Y or X.X = Z.Z
as the appropriate constraint there. If it generated the latter,
it'd erroneously suppose that the Z scan would take care of matters.
To fix, force X.X = Y.Y to be generated and applied at that join node
when this case occurs.
This is *extremely* hard to hit in practice, because various planner
behaviors conspire to mask the problem; starting with the fact that the
planner doesn't really like to generate a parameterized plan of the
above shape. (It might have been impossible to hit it before we
tweaked things to allow this plan shape for star-schema cases.) Many
thanks to Alexander Kirkouski for submitting a reproducible test case.
The bug can be demonstrated in all branches back to 9.2 where parameterized
paths were introduced, so back-patch that far.
Without a few entries beyond old_snapshot_threshold, the lookup
would often fail, resulting in the more aggressive pruning or
vacuum being skipped often enough to matter. This was very clearly
shown by a python test script posted by Ants Aasma, and was likely
a factor in an earlier but somewhat less clear-cut test case posted
by Jeff Janes.
This patch makes no change to the logic, per se -- it just makes
the array of mapping entries big enough to make lookup misses based
on timing much less likely. An occasional miss is still possible
if a thread stalls for more than 10 minutes, but that does not
create any problem with correctness of behavior. Besides, if
things are so busy that a thread is stalling for more than 10
minutes, it is probably OK to skip the more aggressive cleanup at
that particular point in time.
Adjust the way we detect the locale. As a result the minumum Windows
version supported by VS2015 and later is Windows Vista. Add some tweaks
to remove new compiler warnings. Remove documentation references to the
now obsolete msysGit.
Michael Paquier, somewhat edited by me, reviewed by Christian Ullrich.
Backpatch to 9.5
Commit 23a41573c attempted to fix the DatumGetBool macro to ignore bits
in a Datum that are to the left of the actual bool value. But it did that
by casting the Datum to bool; and on compilers that use C99 semantics for
bool, that ends up being a whole-word test, not a 1-byte test. This seems
to be the true explanation for contrib/seg failing in VS2015. To fix, use
GET_1_BYTE() explicitly. I think in the previous patch, I'd had some idea
of not having to commit to bool being exactly 1 byte wide, but regardless
of what the compiler's bool is, boolean columns and Datums are certainly
1 byte wide.
The previous fix was (eventually) back-patched into all active versions,
so do likewise with this one.
Previously, ginInsertCleanup could exit early if it detects that someone else
is cleaning up the pending list, without waiting for that someone else to
finish the job. But in this case vacuum could miss tuples to be deleted.
Cleanup process now locks metapage with a help of heavyweight
LockPage(ExclusiveLock), and it guarantees that there is no another cleanup
process at the same time. Lock is taken differently depending on caller of
cleanup process: any vacuums and gin_clean_pending_list() will be blocked
until lock becomes available, ordinary insert uses conditional lock to
prevent indefinite waiting on lock.
Insert into pending list doesn't use this lock, so insertion isn't blocked.
Also, patch adds stopping of cleanup process when at-start-cleanup-tail is
reached in order to prevent infinite cleanup in case of massive insertion. But
it will stop only for automatic maintenance tasks like autovacuum.
Patch introduces choice of limit of memory to use: autovacuum_work_mem,
maintenance_work_mem or work_mem depending on call path.
Patch for previous releases should be reworked due to changes between 9.6 and
previous ones in this area.
Discover and diagnostics by Jeff Janes and Tomas Vondra
Patch by me with some ideas of Jeff Janes
Commit 989be0810d added a flex/bison lexer/parser to interpret
synchronous_standby_names. It was done in a pretty crufty way, though,
making assorted end-use sites responsible for calling the parser at the
right times. That was not only vulnerable to errors of omission, but made
it possible that lexer/parser errors occur at very undesirable times,
and created memory leakages even if there was no error.
Instead, perform the parsing once during check_synchronous_standby_names
and let guc.c manage the resulting data. To do that, we have to flatten
the parsed representation into a single hunk of malloc'd memory, but that
is not very hard.
While at it, work a little harder on making useful error reports for
parsing problems; the previous code felt that "synchronous_standby_names
parser returned 1" was an appropriate user-facing error message. (To
be fair, it did also log a syntax error message, but separately from the
GUC problem report, which is at best confusing.) It had some outright
bugs in the face of invalid input, too.
I (tgl) also concluded that we need to restrict unquoted names in
synchronous_standby_names to be just SQL identifiers. The previous coding
would accept darn near anything, which (1) makes the quoting convention
both nearly-unnecessary and formally ambiguous, (2) makes it very hard to
understand what is a syntax error and what is a creative interpretation of
the input as a standby name, and (3) makes it impossible to further extend
the syntax in future without a compatibility break. I presume that we're
intending future extensions of the syntax, else this parsing infrastructure
is massive overkill, so (3) is an important objection. Since we've taken
a compatibility hit for non-identifier names with this change anyway, we
might as well lock things down now and insist that users use double quotes
for standby names that aren't identifiers.
Kyotaro Horiguchi and Tom Lane
The way that PartialAggregate and FinalizeAggregate plan nodes were
displaying output columns before was bogus. Now, FinalizeAggregate
produces the same outputs as an Aggregate would have produced, while
PartialAggregate produces each of those outputs prefixed by the word
PARTIAL.
Discussion: 12585.1460737650@sss.pgh.pa.us
Patch by me, reviewed by David Rowley.
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
pg_atomic_compare_exchange_*_impl() were providing only the semantics of
an acquire barrier. Buildfarm members hornet and mandrill revealed this
deficit beginning with commit 008608b9d5.
While we have no report of symptoms in 9.5, we can't rule out the
possibility of certain compilers, hardware, or extension code relying on
these functions' specified barrier semantics. Back-patch to 9.5, where
commit b64d92f1a5 introduced atomics.
Reviewed by Andres Freund.
The coverage was rather lean for cases that bind() or listen() might
return. Add entries for everything that there's a direct equivalent
for in the set of Unix errnos that elog.c has heard of.
There's no longer a need for the MSVC-version-specific code stanza that
forcibly redefines errno code symbols, because since commit 73838b52 we're
unconditionally redefining them in the stanza before this one anyway.
Now it's merely confusing and ugly, so get rid of it; and improve the
comment that explains what's going on here.
Although this is just cosmetic, back-patch anyway since I'm intending
to back-patch some less-cosmetic changes in this same hunk of code.
Revert commit 7cb1db1d95, which represented
a misunderstanding of the problem (if snapmgr.h weren't already included
in bufmgr.h, things wouldn't compile anywhere). Instead install what
I think is the real fix.
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.
Commit 36a35c550a turned the interface between ginPlaceToPage and
its subroutines in gindatapage.c and ginentrypage.c into a royal mess:
page-update critical sections were started in one place and finished in
another place not even in the same file, and the very same subroutine
might return having started a critical section or not. Subsequent patches
band-aided over some of the problems with this design by making things
even messier.
One user-visible resulting problem is memory leaks caused by the need for
the subroutines to allocate storage that would survive until ginPlaceToPage
calls XLogInsert (as reported by Julien Rouhaud). This would not typically
be noticeable during retail index updates. It could be visible in a GIN
index build, in the form of memory consumption swelling to several times
the commanded maintenance_work_mem.
Another rather nasty problem is that in the internal-page-splitting code
path, we would clear the child page's GIN_INCOMPLETE_SPLIT flag well before
entering the critical section that it's supposed to be cleared in; a
failure in between would leave the index in a corrupt state. There were
also assorted coding-rule violations with little immediate consequence but
possible long-term hazards, such as beginning an XLogInsert sequence before
entering a critical section, or calling elog(DEBUG) inside a critical
section.
To fix, redefine the API between ginPlaceToPage() and its subroutines
by splitting the subroutines into two parts. The "beginPlaceToPage"
subroutine does what can be done outside a critical section, including
full computation of the result pages into temporary storage when we're
going to split the target page. The "execPlaceToPage" subroutine is called
within a critical section established by ginPlaceToPage(), and it handles
the actual page update in the non-split code path. The critical section,
as well as the XLOG insertion call sequence, are both now always started
and finished in ginPlaceToPage(). Also, make ginPlaceToPage() create and
work in a short-lived memory context to eliminate the leakage problem.
(Since a short-lived memory context had been getting created in the most
common code path in the subroutines, this shouldn't cause any noticeable
performance penalty; we're just moving the overhead up one call level.)
In passing, fix a bunch of comments that had gone unmaintained throughout
all this klugery.
Report: <571276DD.5050303@dalibo.com>
Robert Haas