Before 728bd991c3, that has improved the support for 2PC files during
recovery, the initial logic scanning files in pg_twophase was done so as
files in the future of the transaction ID horizon were checked first,
followed by a check if a transaction ID is aborted or committed which
could involve a pg_xact lookup. After this commit, these checks have
been done in reverse order.
Files detected as in the future do not have a state that can be checked
in pg_xact, hence this caused recovery to fail abruptly should an
orphaned 2PC file in the future of the transaction ID horizon exist in
pg_twophase at the beginning of recovery.
A test is added to check for this scenario, using an empty 2PC with a
transaction ID large enough to be in the future when running the test.
This test is added in 16 and older versions for now. 17 and newer
versions are impacted by a second bug caused by the addition of the
epoch in the 2PC file names. An equivalent test will be added in these
branches in a follow-up commit, once the second set of issues reported
are fixed.
Author: Vitaly Davydov, Michael Paquier
Discussion: https://postgr.es/m/11e597-676ab680-8d-374f23c0@145466129
Backpatch-through: 13
These arrays were sized with Natts_pg_trigger (19) when they should have
been sized with Natts_pg_event_trigger (7). We'd better fix this as
it's clearly a mistake and it could become problematic if
pg_event_trigger were to gain a dozen or so more columns in the future.
No backpatch as there's no actual bug and the column count on those
tables isn't going to change in released versions.
Author: Xin Zhang <zhanghien@qq.com>
Discussion: https://postgr.es/m/tencent_05AD0FB321A414EC3661204D2102AA6EF605@qq.com
Commit 34486b609 effectively redefined isBackgroundWorker as meaning
"not a regular backend", whereas before it had the narrower
meaning of AmBackgroundWorkerProcess(). For clarity, rename the
field to isRegularBackend and invert its sense.
Discussion: https://postgr.es/m/1808397.1735156190@sss.pgh.pa.us
Cause parallel workers to not check datallowconn, rolcanlogin, and
ACL_CONNECT privileges. The leader already checked these things
(except for rolcanlogin which might have been checked for a different
role). Re-checking can accomplish little except to induce unexpected
failures in applications that might not even be aware that their query
has been parallelized. We already had the principle that parallel
workers rely on their leader to pass a valid set of authorization
information, so this change just extends that a bit further.
Also, modify the ReservedConnections, datconnlimit and rolconnlimit
logic so that these limits are only enforced against regular backends,
and only regular backends are counted while checking if the limits
were already reached. Previously, background processes that had an
assigned database or role were subject to these limits (with rather
random exclusions for autovac workers and walsenders), and the set of
existing processes that counted against each limit was quite haphazard
as well. The point of these limits, AFAICS, is to ensure the
availability of PGPROC slots for regular backends. Since all other
types of processes have their own separate pools of PGPROC slots, it
makes no sense either to enforce these limits against them or to count
them while enforcing the limit.
While edge-case failures of these sorts have been possible for a
long time, the problem got a good deal worse with commit 5a2fed911
(CVE-2024-10978), which caused parallel workers to make some of these
checks using the leader's current role where before we had used its
AuthenticatedUserId, thus allowing parallel queries to fail after
SET ROLE. The previous behavior was fairly accidental and I have
no desire to return to it.
This patch includes reverting 73c9f91a1, which was an emergency hack
to suppress these same checks in some cases. It wasn't complete,
as shown by a recent bug report from Laurenz Albe. We can also revert
fd4d93d26 and 492217301, which hacked around the same problems in one
regression test.
In passing, remove the special case for autovac workers in
CheckMyDatabase; it seems cleaner to have AutoVacWorkerMain pass
the INIT_PG_OVERRIDE_ALLOW_CONNS flag, now that that does what's
needed.
Like 5a2fed911, back-patch to supported branches (which sadly no
longer includes v12).
Discussion: https://postgr.es/m/1808397.1735156190@sss.pgh.pa.us
The need for this was missed in commit 93db6cbda, with the result
being that if we launch a slotsync worker it would consume one of
the PGPROCs in the max_connections pool. That could lead to inability
to launch the worker, or to subsequent failures of connection requests
that should have succeeded according to the configured settings.
Rather than create some one-off infrastructure to support this,
let's group the slotsync worker with the existing autovac launcher
in a new category of "special worker" processes. These are kind of
like auxiliary processes, but they cannot use that infrastructure
because they need to be able to run transactions.
For the moment, make these processes share the PGPROC freelist
used for autovac workers (which previously supplied the autovac
launcher too). This is partly to avoid an ABI change in v17,
and partly because it seems silly to have a freelist with
at most two members. This might be worth revisiting if we grow
enough workers in this category.
Tom Lane and Hou Zhijie. Back-patch to v17.
Discussion: https://postgr.es/m/1808397.1735156190@sss.pgh.pa.us
Commit aac2c9b4fd mandated such locking
and attempted to fulfill that mandate, but it missed REASSIGN OWNED.
Hence, it remained possible to lose VACUUM's inplace update of
datfrozenxid if a REASSIGN OWNED processed that database at the same
time. This didn't affect the other inplace-updated catalog, pg_class.
For pg_class, REASSIGN OWNED calls ATExecChangeOwner() instead of the
generic AlterObjectOwner_internal(), and ATExecChangeOwner() fulfills
the locking mandate.
Like in GRANT, implement this by following the locking protocol for any
catalog subject to the generic AlterObjectOwner_internal(). It would
suffice to do this for IsInplaceUpdateOid() catalogs only. Back-patch
to v13 (all supported versions).
Kirill Reshke. Reported by Alexander Kukushkin.
Discussion: https://postgr.es/m/CAFh8B=mpKjAy4Cuun-HP-f_vRzh2HSvYFG3rhVfYbfEBUhBAGg@mail.gmail.com
This adjusts slot_deform_heap_tuple() to add special-case loops to
eliminate much of the branching that was done within the body of the
main deform loop.
Previously, while looping over each attribute to deform,
slot_deform_heap_tuple() would always recheck if the given attribute was
NULL by looking at HeapTupleHasNulls() and if so, went on to check the
tuple's NULL bitmap. Since many tuples won't contain any NULLs, we can
just check HeapTupleHasNulls() once and when there are no NULLs, use a
more compact version of the deforming loop which contains no NULL checking
code at all.
The same is possible for the "slow" mode checking part of the loop. That
variable was checked several times for each attribute, once to determine
if the offset to the attribute value could be taken from the attcacheoff,
and again to check if the offset could be cached for next time.
These "slow" checks can mostly be eliminated by instead having multiple
loops. Initially, we can start in the non-slow loop and break out of
that loop if and only if we must stop caching the offset. This
eliminates branching for both slow and non-slow deforming methods. The
amount of code required for the no nulls / non-slow version is very
small. It's possible to have separate loops like this due to the fact
that once we move into slow mode, we never need to switch back into
non-slow mode for a given tuple.
We have the compiler take care of writing out the multiple required
loops by having a pg_attribute_always_inline function which gets called
various times passing in constant values for the "slow" and "hasnulls"
parameters. This allows the compiler to eliminate const-false branches
and remove comparisons for const-true ones.
This commit has shown overall query performance increases of around 5-20%
in deform-heavy OLAP-type workloads.
Author: David Rowley
Reviewed-by: Victor Yegorov
Discussion: https://postgr.es/m/CAGnEbog92Og2CpC2S8=g_HozGsWtt_3kRS1sXjLz0jKSoCNfLw@mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvo9e0XG71WrefYaRv5n4xNPLK4k8LjD0mSR3c9KR2vi2Q@mail.gmail.com
Previously, if an infinite value was passed to date_trunc(), then the
same infinite value would always be returned regardless of the field
unit given by the caller. This commit updates the function so that an
error is returned when an invalid unit is passed to date_trunc() with an
infinite value.
This matches the behavior of date_trunc() with a finite value and
date_part() with an infinite value, making the handling of interval,
timestamp and timestamptz more consistent across the board for these two
functions.
Some tests are added to cover all these new failure cases, with an
unsupported unit and infinite values for the three data types. There
were no test cases in core that checked all these patterns up to now.
Author: Joseph Koshakow
Discussion: https://postgr.es/m/CAAvxfHc4084dGzEJR0_pBZkDuqbPGc5wn7gK_M0XR_kRiCdUJQ@mail.gmail.com
The totalrows parameter in compute_expr_stats is unused, so remove it.
This is a static function, so the parameter can easily be added again if
it's ever needed.
Author: Ilia Evdokimov <ilya.evdokimov@tantorlabs.ru>
Discussion: https://postgr.es/m/667b92d2-f953-4fcb-9377-3765f5b94187@tantorlabs.com
Rename "core_yy_extra_type core_yy" to "core_yy_extra". The previous
name was a bit unclear and confusing. The new name matches the name
used elsewhere for the same purpose, for example in
src/backend/parser/gramparse.h.
This fixes "unresolved external symbol" errors with extensions that
use functions from libpgport that need special CFLAGS to
compile. Currently, that includes the CRC-32 functions.
Commit 2571c1d5cc did this for libcommon, but I missed that libpqport
has the same issue.
Reported-by: Tom Lane
Backpatch-through: 16, where Meson was introduced
Discussion: https://www.postgresql.org/message-id/CAOdR5yF0krWrxycA04rgUKCgKugRvGWzzGLAhDZ9bzNv8g0Lag@mail.gmail.com
Trying to clean up the code a bit while we're working on these files
for the reentrant scanner/pure parser patches. This cleanup only
touches the code sections after the second '%%' in each file, via a
manually-supervised and locally hacked up pgindent.
Use the flex %option reentrant to make the generated scanner
reentrant and thread-safe. Note: The parser was already pure.
Simplify flex scan buffer management: Instead of constructing the
buffer from pieces and then using yy_scan_buffer(), we can just use
yy_scan_string(), which does the same thing internally. (Actually, we
use yy_scan_bytes() here because we already have the length.)
Use flex yyextra to handle context information, instead of global
variables. This complements the other changes to make the scanner
reentrant.
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://www.postgresql.org/message-id/flat/eb6faeac-2a8a-4b69-9189-c33c520e5b7b@eisentraut.org
Use the flex %option reentrant and the bison option %pure-parser to
make the generated scanner and parser pure, reentrant, and
thread-safe.
Make the generated scanner use palloc() etc. instead of malloc() etc.
Previously, we only used palloc() for the buffer, but flex would still
use malloc() for its internal structures. Now, all the memory is
under palloc() control.
Simplify flex scan buffer management: Instead of constructing the
buffer from pieces and then using yy_scan_buffer(), we can just use
yy_scan_string(), which does the same thing internally.
The previous code was necessary because we allocated the buffer with
palloc() and the rest of the state was handled by malloc(). But this
is no longer the case; everything is under palloc() now.
Use flex yyextra to handle context information, instead of global
variables. This complements the other changes to make the scanner
reentrant.
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://www.postgresql.org/message-id/flat/eb6faeac-2a8a-4b69-9189-c33c520e5b7b@eisentraut.org
Use the flex %option reentrant and the bison option %pure-parser to
make the generated scanner and parser pure, reentrant, and
thread-safe.
Make the generated scanner use palloc() etc. instead of malloc() etc.
Previously, we only used palloc() for the buffer, but flex would still
use malloc() for its internal structures. As a result, there could be
some small memory leaks in case of uncaught errors. Now, all the
memory is under palloc() control, so there are no more such issues.
Simplify flex scan buffer management: Instead of constructing the
buffer from pieces and then using yy_scan_buffer(), we can just use
yy_scan_string(), which does the same thing internally.
The previous code was necessary because we allocated the buffer with
palloc() and the rest of the state was handled by malloc(). But this
is no longer the case; everything is under palloc() now.
Use flex yyextra to handle context information, instead of global
variables. This complements the other changes to make the scanner
reentrant.
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Co-authored-by: Andreas Karlsson <andreas@proxel.se>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://www.postgresql.org/message-id/flat/eb6faeac-2a8a-4b69-9189-c33c520e5b7b@eisentraut.org
Per git log, the last time someone tried to do something with
pgrminclude was around 2011. Many (not all) of the "pgrminclude
ignore" annotations are of a newer date but seem to have just been
copied around during refactorings and file moves and don't seem to
reflect an actual need anymore.
There have been some parallel experiments with include-what-you-use
(IWYU) annotations, but these don't seem to correspond very strongly
to pgrminclude annotations, so there is no value in keeping the
existing ones even for that kind of thing.
So, wipe them all away. We can always add new ones in the future
based on actual needs.
Discussion: https://www.postgresql.org/message-id/flat/2d4dc7b2-cb2e-49b1-b8ca-ba5f7024f05b%40eisentraut.org
5983a4cff added CompactAttribute as an abbreviated alternative to
FormData_pg_attribute to allow more cache-friendly processing in tasks
related to TupleDescs. That commit contained some assert-only code to
check that the CompactAttribute had been populated correctly, however,
the method used to do that checking caused the TupleDesc's
CompactAttribute to be zeroed before it was repopulated and compared to
the snapshot taken before the memset call. This caused issues as the type
cache caches TupleDescs in shared memory which can be used by multiple
backend processes at the same time. There was a window of time between
the zero and repopulation of the CompactAttribute where another process
would mistakenly think that the CompactAttribute is invalid due to the
memset.
To fix this, instead of taking a snapshot of the CompactAttribute and
calling populate_compact_attribute() and comparing the snapshot to the
freshly populated TupleDesc's CompactAttribute, refactor things so we
can just populate a temporary CompactAttribute on the stack. This way
we don't touch the TupleDesc's memory.
Reported-by: Alexander Lakhin, SQLsmith
Discussion: https://postgr.es/m/ca3a256a-5d12-42db-aabe-a75a030d9fb9@gmail.com
These two platforms have a remarkably tight default limit on the
number of SysV semaphores in the system: SEMMNS is only 60
out-of-the-box. Unless manual action is taken to raise that,
we'll only be able to allocate 3 sets of 16 usable semaphores
each, leading to initdb setting max_connections to just 20.
That's problematic because the core regression tests expect
to be able to launch 20 concurrent sessions, leaving us with
no headroom. This seems to be the cause of intermittent
buildfarm failures on some machines.
While there's no getting around the fact that you'd better raise
SEMMNS for production use on these platforms, it does seem desirable
for "make check" to pass reliably without that. We can make that
happen, at least for awhile longer, with two small changes:
* Change sysv_sema.c's SEMAS_PER_SET to 19, so that we can eat up
all of the available semas not just most of them.
* Change initdb to make the smallest max_connections value it will
consider be 25 not 20.
As of HEAD this will leave us with four free semaphores (using the
default values for other relevant parameters such as max_wal_senders).
So we won't need to consider this again until we've invented five
more background processes. Maybe by then we can switch both these
platforms to some other semaphore API.
For the moment, do this only in master; there've not been field
complaints that might justify a back-patch.
Discussion: https://postgr.es/m/db2773a2-aca0-43d0-99c1-060efcd9954e@gmail.com
Reset btpo_cycleid to 0 in btree_xlog_vacuum for consistency with
_bt_delitems_vacuum (the corresponding original execution code). This
makes things neater.
There might be some performance benefit to being consistent like this.
When btvacuumpage doesn't call _bt_delitems_vacuum, it can still
proactively reset btpo_cycleid to 0 via a separate hint-like update
mechanism (it does so whenever it sees that it isn't already set to 0).
And so it's possible that being consistent about resetting btpo_cycleid
like this will save work later on, after standby promotion: subsequent
VACUUMs won't need to clear btpo_cycleid using the hint-like update
mechanism as often as they otherwise would.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Andrey Borodin <x4mmm@yandex-team.ru>
Discussion: https://postgr.es/m/CAH2-Wz=+LDFxn9NZyEsCo8ifcyKt6+n-VLyygySEHgMz+oynqw@mail.gmail.com
Despite the best efforts of commit 0e5c82380, we're still seeing
occasional failures of postgres_fdw's query_cancel test in the
buildfarm. Investigation suggests that its 100ms timeout is
still not enough to reliably ensure that the remote side starts
the query before receiving the cancel request --- and if it
hasn't, it will just discard the request because it's idle.
We discussed allowing a cancel request to kill the next-received
query, but that would have wide and perhaps unpleasant side-effects.
What seems safer is to make postgres_fdw do what a human user would
likely do, which is issue another cancel request if the first one
didn't seem to do anything. We'll keep the same overall 30 second
grace period before concluding things are broken, but issue additional
cancel requests after 1 second, then 2 more seconds, then 4, then 8.
(The next one in series is 16 seconds, but we'll hit the 30 second
timeout before that.)
Having done that, revert the timeout in query_cancel.sql to 10 ms.
That will still be enough on most machines, most of the time, for
the remote query to start; but now we're intentionally risking the
race condition occurring sometimes in the buildfarm, so that the
repeat-cancel code path will get some testing.
As before, back-patch to v17. We might eventually contemplate
back-patching this further, and/or adding similar logic to dblink.
But given the lack of field complaints to date, this feels like
mostly an exercise in test case stabilization, so v17 is enough.
Discussion: https://postgr.es/m/colnv3lzzmc53iu5qoawynr6qq7etn47lmggqr65ddtpjliq5d@glkveb4m6nop
A historic snapshot should only be used for catalog access, not
general queries. We never call GetTransactionSnapshot() during logical
decoding, which is good because it wouldn't be very sensible, so the
code to deal with that was unreachable and untested. Turn it into an
error, to avoid doing that in the future either.
Discussion: https://www.postgresql.org/message-id/a868fe78-ddb4-4b0a-9b96-873d91d93cfd@iki.fi
In get_database_list() and get_subscription_list(), the
GetTransactionSnapshot() call is not required because the catalog
table scans use the catalog snapshot, which is held until the end of
the scan. See table_beginscan_catalog(), which calls
RegisterSnapshot(GetCatalogSnapshot(relid)).
In InitPostgres, it's a little less obvious that it's not required,
but still true I believe. All the catalog lookups in InitPostgres()
also use the catalog snapshot, and the looked up values are copied
while still holding the snapshot.
Furthermore, as the removed FIXME comments said, calling
GetTransactionSnapshot() didn't really prevent MyProc->xmin from being
reset anyway.
Discussion: https://www.postgresql.org/message-id/7c56f180-b9e1-481e-8c1d-efa63de3ecbb@iki.fi
Jian He reported the src/include/utility/tcop.h one and the remainder
were found by using a script to look for src/* and check that we have a
filename or directory of that name.
In passing, fix some out-date comments.
Reported-by: Jian He <jian.universality@gmail.com>
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/CACJufxGoE3H-7VgO02=PrR4SNuVWDVbfTyUnwO0HvS-Lxurnog@mail.gmail.com
Library unloading has never been supported with its code removed in
ab02d702ef, and there were some comments still mentioning that it was
a possible operation.
ChangAo has noticed the incorrect references in dfmgr.c, while I have
noticed the other ones while scanning the rest of the tree for similar
mistakes.
Author: ChangAo Chen, Michael Paquier
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/tencent_1D09840A1632D406A610C8C4E2491D74DB0A@qq.com
GetSnapshotData() set TransactionXmin = MyProc->xmin, but when
SnapshotResetXmin() advanced MyProc->xmin, it did not advance
TransactionXmin correspondingly. That meant that TransactionXmin could
be older than MyProc->xmin, and XIDs between than TransactionXmin and
the real MyProc->xmin could be vacuumed away. One known consequence is
in pg_subtrans lookups: we might try to look up the status of an XID
that was already truncated away.
Back-patch to all supported versions.
Reviewed-by: Andres Freund
Discussion: https://www.postgresql.org/message-id/d27a046d-a1e4-47d1-a95c-fbabe41debb4@iki.fi
Here we convert CompactAttribute.attalign from a char, which is directly
derived from pg_attribute.attalign into a uint8, which stores the number
of bytes to align the column's value by in the tuple.
This allows tuple deformation and tuple size calculations to move away
from using the inefficient att_align_nominal() macro, which manually
checks each TYPALIGN_* char to translate that into the alignment bytes
for the given type. Effectively, this commit changes those to TYPEALIGN
calls, which are branchless and only perform some simple arithmetic with
some bit-twiddling.
The removed branches were often mispredicted by CPUs, especially so in
real-world tables which often contain a mishmash of different types
with different alignment requirements.
Author: David Rowley
Reviewed-by: Andres Freund, Victor Yegorov
Discussion: https://postgr.es/m/CAApHDvrBztXP3yx=NKNmo3xwFAFhEdyPnvrDg3=M0RhDs+4vYw@mail.gmail.com
This used to say "nsubxcnt isn't decreased when subtransactions
abort", but there's no variable called nsubxcnt. Commit 8548ddc61b
changed it to "subxcnt", among other typo fixes, but that was wrong
too: the comment actually talks about txn->nsubtxns. That's the field
that's incremented but never decremented and is used for the
allocation earlier in the function.
RelationTruncate() does three things, while holding an
AccessExclusiveLock and preventing checkpoints:
1. Logs the truncation.
2. Drops buffers, even if they're dirty.
3. Truncates some number of files.
Step 2 could previously be canceled if it had to wait for I/O, and step
3 could and still can fail in file APIs. All orderings of these
operations have data corruption hazards if interrupted, so we can't give
up until the whole operation is done. When dirty pages were discarded
but the corresponding blocks were left on disk due to ERROR, old page
versions could come back from disk, reviving deleted data (see
pgsql-bugs #18146 and several like it). When primary and standby were
allowed to disagree on relation size, standbys could panic (see
pgsql-bugs #18426) or revive data unknown to visibility management on
the primary (theorized).
Changes:
* WAL is now unconditionally flushed first
* smgrtruncate() is now called in a critical section, preventing
interrupts and causing PANIC on file API failure
* smgrtruncate() has a new parameter for existing fork sizes,
because it can't call smgrnblocks() itself inside a critical section
The changes apply to RelationTruncate(), smgr_redo() and
pg_truncate_visibility_map(). That last is also brought up to date with
other evolutions of the truncation protocol.
The VACUUM FileTruncate() failure mode had been discussed in older
reports than the ones referenced below, with independent analysis from
many people, but earlier theories on how to fix it were too complicated
to back-patch. The more recently invented cancellation bug was
diagnosed by Alexander Lakhin. Other corruption scenarios were spotted
by me while iterating on this patch and earlier commit 75818b3a.
Back-patch to all supported releases.
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reported-by: rootcause000@gmail.com
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/18146-04e908c662113ad5%40postgresql.org
Discussion: https://postgr.es/m/18426-2d18da6586f152d6%40postgresql.org
One test added in 9aea73fc61 did not take into account that the
backend may have some fsync even after a checkpoint. Let's relax it to
be more flexible.
Per report from buildfarm member grassquit, via Alexander Lakhin.
Author: Bertrand Drouvot
Discussion: https://postgr.es/m/6143ab0a-9e88-4790-8d9d-50ba45657761@gmail.com
The new compact_attrs array stores a few select fields from
FormData_pg_attribute in a more compact way, using only 16 bytes per
column instead of the 104 bytes that FormData_pg_attribute uses. Using
CompactAttribute allows performance-critical operations such as tuple
deformation to be performed without looking at the FormData_pg_attribute
element in TupleDesc which means fewer cacheline accesses.
For some workloads, tuple deformation can be the most CPU intensive part
of processing the query. Some testing with 16 columns on a table
where the first column is variable length showed around a 10% increase in
transactions per second for an OLAP type query performing aggregation on
the 16th column. However, in certain cases, the increases were much
higher, up to ~25% on one AMD Zen4 machine.
This also makes pg_attribute.attcacheoff redundant. A follow-on commit
will remove it, thus shrinking the FormData_pg_attribute struct by 4
bytes.
Author: David Rowley
Reviewed-by: Andres Freund, Victor Yegorov
Discussion: https://postgr.es/m/CAApHDvrBztXP3yx=NKNmo3xwFAFhEdyPnvrDg3=M0RhDs+4vYw@mail.gmail.com
b7493e1ab3 removed leftover mentions of XLOG_HEAP2_FREEZE_PAGE records
from comments but neglected to remove one mention of FREEZE_PAGE.
Reported off-list by Alexander Lakhin
It was reasonable to preserve the old API of BuildTupleHashTable()
in the back branches, but in HEAD we should actively discourage use
of that version. There are no remaining callers in core, so just
get rid of it. Then rename BuildTupleHashTableExt() back to
BuildTupleHashTable().
While at it, fix up the miserably-poorly-maintained header comment
for BuildTupleHashTable[Ext]. It looks like more than one patch in
this area has had the opinion that updating comments is beneath them.
Discussion: https://postgr.es/m/538343.1734646986@sss.pgh.pa.us
Append, MergeAppend, and RecursiveUnion can all use the support
functions added in commit 276279295. The first two can report a
fixed result slot type if all their children return the same fixed
slot type. That does nothing for the append step itself, but might
allow optimizations in the parent plan node. RecursiveUnion can
optimize tuple hash table operations in the same way as SetOp now
does.
Patch by me; thanks to Richard Guo and David Rowley for review.
Discussion: https://postgr.es/m/1850138.1731549611@sss.pgh.pa.us
Remove the code for inserting flag columns in the inputs of a SetOp.
That was the only reason why there would be resjunk columns in a
set-operations plan tree, so we can get rid of some code that
supported that, too.
Get rid of choose_hashed_setop() in favor of building Paths for
the hashed and sorted alternatives, and letting them fight it out
within add_path().
Remove set_operation_ordered_results_useful(), which was giving wrong
answers due to examining the wrong ancestor node: we need to examine
the immediate SetOperationStmt parent not the topmost node. Instead
make each caller of recurse_set_operations() pass down the relevant
parent node. (This thinko seems to have led only to wasted planning
cycles and possibly-inferior plans, not wrong query answers. Perhaps
we should back-patch it, but I'm not doing so right now.)
Teach generate_nonunion_paths() to consider pre-sorted inputs for
sorted SetOps, rather than always generating a Sort node.
Patch by me; thanks to Richard Guo and David Rowley for review.
Discussion: https://postgr.es/m/1850138.1731549611@sss.pgh.pa.us
The original design for set operations involved appending the two
input relations into one and adding a flag column that allows
distinguishing which side each row came from. Then the SetOp node
pries them apart again based on the flag. This is bizarre. The
only apparent reason to do it is that when sorting, we'd only need
one Sort node not two. But since sorting is at least O(N log N),
sorting all the data is actually worse than sorting each side
separately --- plus, we have no chance of taking advantage of
presorted input. On top of that, adding the flag column frequently
requires an additional projection step that adds cycles, and then
the Append node isn't free either. Let's get rid of all of that
and make the SetOp node have two separate children, using the
existing outerPlan/innerPlan infrastructure.
This initial patch re-implements nodeSetop.c and does a bare minimum
of work on the planner side to generate correctly-shaped plans.
In particular, I've tried not to change the cost estimates here,
so that the visible changes in the regression test results will only
involve removal of useless projection steps and not any changes in
whether to use sorted vs hashed mode.
For SORTED mode, we combine successive identical tuples from each
input into groups, and then merge-join the groups. The tuple
comparisons now use SortSupport instead of simple equality, but
the group-formation part should involve roughly the same number of
tuple comparisons as before. The cross-comparisons between left and
right groups probably add to that, but I'm not sure to quantify how
many more comparisons we might need.
For HASHED mode, nodeSetop's logic is almost the same as before,
just refactored into two separate loops instead of one loop that
has an assumption that it will see all the left-hand inputs first.
In both modes, I added early-exit logic to not bother reading the
right-hand relation if the left-hand input is empty, since neither
INTERSECT nor EXCEPT modes can produce any output if the left input
is empty. This could have been done before in the hashed mode, but
not in sorted mode. Sorted mode can also stop as soon as it exhausts
the left input; any remaining right-hand tuples cannot have matches.
Also, this patch adds some infrastructure for detecting whether
child plan nodes all output the same type of tuple table slot.
If they do, the hash table logic can use slightly more efficient
code based on assuming that that's the input slot type it will see.
We'll make use of that infrastructure in other plan node types later.
Patch by me; thanks to Richard Guo and David Rowley for review.
Discussion: https://postgr.es/m/1850138.1731549611@sss.pgh.pa.us
1a0da347a7 replaced Bitmap Table Scan's separate private and
shared bitmap iterators with a unified iterator. It accidentally set up
the prefetch iterator twice for non-parallel bitmap table scans. Remove
the extra set up call to tbm_begin_iterate().
1a0da347a7 replaced Bitmap Table Scan's individual private and
shared iterators with a unified iterator. It neglected, however, to
check if the iterator had already been cleaned up before doing so on
rescan. Add this check both on rescan and end scan to be safe.
Reported-by: Richard Guo
Author: Richard Guo
Discussion: https://postgr.es/m/CAMbWs48nrhcLY1kcd-u9oD%2B6yiS631F_8Fx8ZGsO-BYDwH%2Bbyw%40mail.gmail.com
Consistently reset so->scanBehind at the beginning of nbtree array
advancement, even during sktrig_required=false calls (calls where array
advancement is triggered by an unsatisfied non-required array scan key).
Otherwise, it's possible for queries to fail to return all relevant
tuples to the scan given a low-order required scan key that was
previously deemed "satisfied" by a truncated high key attribute value.
This only happened at the point where a later non-required array scan
key needed to be "advanced" once on the next leaf page (that is, once
the right sibling of the truncated high key page was reached).
The underlying issue was that later code within _bt_advance_array_keys
assumed that the so->scanBehind flag must have been set using the
current page's high key (not the previous page's high key). Any later
successful recheck call to _bt_check_compare would therefore spuriously
be prevented from making _bt_advance_array_keys return true, based on
the faulty belief that the truncated attribute must be from the scan's
current tuple (i.e. the non-pivot tuple at the start of the next page).
_bt_advance_array_keys would return false for the tuple, ultimately
resulting in _bt_checkkeys failing to return a matching tuple.
Oversight in commit 5bf748b8, which enhanced nbtree ScalarArrayOp
execution.
Author: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/CAH2-WzkJKncfqyAUTeuB5GgRhT1vhsWO2q11dbZNqKmvjopP_g@mail.gmail.com
Backpatch: 17-, where commit 5bf748b8 first appears.
Michael Paquier