READ ONLY transactions should prevent modifications to foreign data as
well as local data, but postgres_fdw transactions declared as READ ONLY
that reference foreign tables mapped to a remote view executing volatile
functions would modify data on remote servers, as it would open remote
transactions in READ WRITE mode.
Similarly, DEFERRABLE transactions should not abort due to a
serialization failure even when accessing foreign data, but postgres_fdw
transactions declared as DEFERRABLE would abort due to that failure in a
remote server, as it would open remote transactions in NOT DEFERRABLE
mode.
To fix, modify postgres_fdw to open remote transactions in the same
access/deferrable modes as the local transaction. This commit also
modifies it to open remote subtransactions in the same access mode as
the local subtransaction.
This commit changes the behavior of READ ONLY/DEFERRABLE transactions
using postgres_fdw; in particular, it doesn't allow the READ ONLY
transactions to modify data on remote servers anymore, so such
transactions should be redeclared as READ WRITE or rewritten using other
tools like dblink. The release notes should note this as an
incompatibility.
These issues exist since the introduction of postgres_fdw, but to avoid
the incompatibility in the back branches, fix them in master only.
Author: Etsuro Fujita <etsuro.fujita@gmail.com>
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/CAPmGK16n_hcUUWuOdmeUS%2Bw4Q6dZvTEDHb%3DOP%3D5JBzo-M3QmpQ%40mail.gmail.com
Discussion: https://postgr.es/m/E1uLe9X-000zsY-2g%40gemulon.postgresql.org
This avoids increasing the distance to the maximum in cases where the I/O
subsystem is already keeping up. This turns out to be important for
performance for two reasons:
- Pinning a lot of buffers is not cheap. If additional pins allow us to avoid
IO waits, it's definitely worth it, but if we can already do all the
necessary readahead at a distance of 16, reading ahead 512 buffers can
increase the CPU overhead substantially. This is particularly noticeable
when the to-be-read blocks are already in the kernel page cache.
- If the read stream is read to completion, reading in data earlier than
needed is of limited consequences, leaving aside the CPU costs mentioned
above. But if the read stream will not be fully consumed, e.g. because it is
on the inner side of a nested loop join, the additional IO can be a serious
performance issue. This is not that commonly a problem for current read
stream users, but the upcoming work, to use a read stream to fetch table
pages as part of an index scan, frequently encounters this.
Note that this commit would have substantial performance downsides without
earlier commits:
- Commit 6e36930f9a, which avoids decreasing the readahead distance when
there was recent IO, is crucial, as otherwise we very often would end up not
reading ahead aggressively enough anymore with this commit, due to
increasing the distance less often.
- "read stream: Split decision about look ahead for AIO and combining" is
important as we would otherwise not perform IO combining when the IO
subsystem can keep up.
- "aio: io_uring: Trigger async processing for large IOs" is important to
continue to benefit from memory copy parallelism when using fewer IOs.
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Tested-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/f3xxfrkafjxpyqxywcxricxgyizjirfceychyxsgn7bwjp5eda@kwbduhy7tfmu
Discussion: https://postgr.es/m/CA+hUKGL2PhFyDoqrHefqasOnaXhSg48t1phs3VM8BAdrZqKZkw@mail.gmail.com
In a subsequent commit the read-ahead distance will only be increased when
waiting for IO. Without further work that would cause a regression: As IO
combining and read-ahead are currently controlled by the same mechanism, we
would end up not allowing IO combining when never needing to wait for IO (as
the distance ends up too small to allow for full sized IOs), which can
increase CPU overhead. A typical reason to not have to wait for IO completion
at a low look-ahead distance is use of io_uring with the to-be-read data in
the page cache. But even with worker the IO submission rate may be low enough
for the worker to keep up.
One might think that we could just always perform IO combining, but doing so
at the start of a scan can cause performance regressions:
1) Performing a large IO commonly has a higher latency than smaller IOs. That
is not a problem once reading ahead far enough, but at the start of a stream
it can lead to longer waits for IO completion.
2) Sometimes read streams will not be read to completion. Immediately starting
with full sized IOs leads to more wasted effort. This is not commonly an
issue with existing read stream users, but the upcoming use of read streams
to fetch table pages as part of an index scan frequently encounters this.
Solve this issue by splitting ReadStream->distance into ->combine_distance and
->readahead_distance. Right now they are increased/decreased at the same time,
but that will change in the next commit.
One of the comments in read_stream_should_look_ahead() refers to a motivation
that only really exists as of the next commit, but without it the code doesn't
make sense on its own.
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/f3xxfrkafjxpyqxywcxricxgyizjirfceychyxsgn7bwjp5eda@kwbduhy7tfmu
Discussion: https://postgr.es/m/CA+hUKGL2PhFyDoqrHefqasOnaXhSg48t1phs3VM8BAdrZqKZkw@mail.gmail.com
The long if statements were hard to read and hard to document. Splitting them
into inline helpers makes it much easier to explain each part separately.
This is done in preparation for making the logic more complicated...
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/f3xxfrkafjxpyqxywcxricxgyizjirfceychyxsgn7bwjp5eda@kwbduhy7tfmu
io_method=io_uring has a heuristic to trigger asynchronous processing of IOs
once the IO depth is a bit larger. That heuristic is important when doing
buffered IO from the kernel page cache, to allow parallelizing of the memory
copy, as otherwise io_method=io_uring would be a lot slower than
io_method=worker in that case.
An upcoming commit will make read_stream.c only increase the read-ahead
distance if we needed to wait for IO to complete. If to-be-read data is in the
kernel page cache, io_uring will synchronously execute IO, unless the IO is
flagged as async. Therefore the aforementioned change in read_stream.c
heuristic would lead to a substantial performance regression with io_uring
when data is in the page cache, as we would never reach a deep enough queue to
actually trigger the existing heuristic.
Parallelizing the copy from the page cache is mainly important when doing a
lot of IO, which commonly is only possible when doing largely sequential IO.
The reason we don't just mark all io_uring IOs as asynchronous is that the
dispatch to a kernel thread has overhead. This overhead is mostly noticeable
with small random IOs with a low queue depth, as in that case the gain from
parallelizing the memory copy is small and the latency cost high.
The facts from the two prior paragraphs show a way out: Use the size of the IO
in addition to the depth of the queue to trigger asynchronous processing.
One might think that just using the IO size might be enough, but
experimentation has shown that not to be the case - with deep look-ahead
distances being able to parallelize the memory copy is important even with
smaller IOs.
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/f3xxfrkafjxpyqxywcxricxgyizjirfceychyxsgn7bwjp5eda@kwbduhy7tfmu
Discussion: https://postgr.es/m/CA+hUKGL2PhFyDoqrHefqasOnaXhSg48t1phs3VM8BAdrZqKZkw@mail.gmail.com
Also rename it to index_create_copy. Add a 'boolean concurrent' option,
and make it work for both cases: in concurrent mode, just create the
catalog entries; caller is responsible for the actual building later.
In non-concurrent mode, the index is built right away.
This allows it to be reused for other purposes -- specifically, for
concurrent REPACK.
(With the CONCURRENTLY option, REPACK cannot simply swap the heap file and
rebuild its indexes. Instead, it needs to build a separate set of
indexes, including their system catalog entries, *before* the actual
swap, to reduce the time AccessExclusiveLock needs to be held for. This
approach is different from what CREATE INDEX CONCURRENTLY does.)
Per a suggestion from Mihail Nikalayeu.
Author: Antonin Houska <ah@cybertec.at>
Reviewed-by: Mihail Nikalayeu <mihailnikalayeu@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Discussion: https://postgr.es/m/41104.1754922120@localhost
Avoid dropping the heap page pin (xs_cbuf) and visibility map pin
(xs_vmbuffer) within heapam_index_fetch_reset. Retaining these pins
saves cycles during certain nested loop joins and merge joins that
frequently restore a saved mark: cases where the next tuple fetched
after a reset often falls on the same heap page will now avoid the cost
of repeated pinning and unpinning.
Avoiding dropping the scan's heap page buffer pin is preparation for an
upcoming patch that will add I/O prefetching to index scans. Testing of
that patch (which makes heapam tend to pin more buffers concurrently
than was typical before now) shows that the aforementioned cases get a
small but clearly measurable benefit from this optimization.
Upcoming work to add a slot-based table AM interface for index scans
(which is further preparation for prefetching) will move VM checks for
index-only scans out of the executor and into heapam. That will expand
the role of xs_vmbuffer to include VM lookups for index-only scans (the
field won't just be used for setting pages all-visible during on-access
pruning via the enhancement recently introduced by commit b46e1e54).
Avoiding dropping the xs_vmbuffer pin will preserve the historical
behavior of nodeIndexonlyscan.c, which always kept this pin on a rescan;
that aspect of this commit isn't really new.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CAH2-Wz=g=JTSyDB4UtB5su2ZcvsS7VbP+ZMvvaG6ABoCb+s8Lw@mail.gmail.com
Add an explicit BlockNumber field (xs_blk) to IndexFetchHeapData that
tracks which heap block is currently pinned in xs_cbuf.
heapam_index_fetch_tuple now uses xs_blk to determine when buffer
switching is needed, replacing the previous approach that compared
buffer identities via ReleaseAndReadBuffer on every non-HOT-chain call.
This is preparatory work for an upcoming commit that will add index
prefetching using a read stream. Delegating the release of a currently
pinned buffer to ReleaseAndReadBuffer won't work anymore -- at least not
when the next buffer that the scan needs to pin is one returned by
read_stream_next_buffer (not a buffer returned by ReadBuffer).
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CAH2-Wz=g=JTSyDB4UtB5su2ZcvsS7VbP+ZMvvaG6ABoCb+s8Lw@mail.gmail.com
Move the heapam index fetch callbacks (index_fetch_begin,
index_fetch_reset, index_fetch_end, and index_fetch_tuple) into a new
dedicated file. Also move heap_hot_search_buffer over. This is a
purely mechanical move with no functional impact.
Upcoming work to add a slot-based table AM interface for index scans
will substantially expand this code. Keeping it in heapam_handler.c
would clutter a file whose primary role is to wire up the TableAmRoutine
callbacks. Bitmap heap scans and sequential scans would benefit from
similar separation in the future.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/bmbrkiyjxoal6o5xadzv5bveoynrt3x37wqch7w3jnwumkq2yo@b4zmtnrfs4mh
Rename heapam_index_fetch_tuple's call_again argument to heap_continue,
for consistency with the pointed-to variable name (IndexScanDescData's
xs_heap_continue field).
Preparation for an upcoming commit that will move index scan related
heapam functions into their own file.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/bmbrkiyjxoal6o5xadzv5bveoynrt3x37wqch7w3jnwumkq2yo@b4zmtnrfs4mh
In similar vein to commit 3c6e8c123, the ARMv8 cryptography extension
has 64x64 -> 128-bit carryless multiplication instructions suitable
for computing CRC. This was tested to be around twice as fast as
scalar CRC instructions for longer inputs.
We now do a runtime check, even for builds that target "armv8-a+crc",
but those builds can still use a direct call for constant inputs,
which we assume are short.
As for x86, the MIT-licensed implementation was generated with the
"generate" program from
https://github.com/corsix/fast-crc32/
Reviewed-by: Nathan Bossart <nathandbossart@gmail.com>
Discussion: https://postgr.es/m/CANWCAZaKhE+RD5KKouUFoxx1EbUNrNhcduM1VQ=DkSDadNEFng@mail.gmail.com
We already rely on autovectorization for computing page checksums,
but on x86 we can get a further several-fold performance increase by
annotating pg_checksum_block() with a function target attribute for
the AVX2 instruction set extension. Not only does that use 256-bit
registers, it can also use vector multiplication rather than the
vector shifts and adds used in SSE2.
Similar to other hardware-specific paths, we set a function pointer
on first use. We don't bother to avoid this on platforms without AVX2
since the overhead of indirect calls doesn't matter for multi-kilobyte
inputs. However, we do arrange so that only core has the function
pointer mechanism. External programs will continue to build a normal
static function and don't need to be aware of this.
This matters most when using io_uring since in that case the checksum
computation is not done in parallel by IO workers.
Co-authored-by: Matthew Sterrett <matthewsterrett2@gmail.com>
Co-authored-by: Andrew Kim <andrew.kim@intel.com>
Reviewed-by: Oleg Tselebrovskiy <o.tselebrovskiy@postgrespro.ru>
Tested-by: Ants Aasma <ants.aasma@cybertec.at>
Tested-by: Stepan Neretin <slpmcf@gmail.com> (earlier version)
Discussion: https://postgr.es/m/CA+vA85_5GTu+HHniSbvvP+8k3=xZO=WE84NPwiKyxztqvpfZ3Q@mail.gmail.com
Discussion: https://postgr.es/m/20250911054220.3784-1-root%40ip-172-31-36-228.ec2.internal
It's been missing ever since fast-path locking was introduced. It's a
small discrepancy, about 4 kB, but let's be tidy. This doesn't seem
worth backpatching, however; in stable branches we were less precise
about the estimates and e.g. added a 10% margin to the hash table
estimates, which is usually much bigger than this discrepancy.
For the three implementations that have caused problems so far:
* GNU and BSD (libarchive) tar both understand --format=ustar
* ustar doesn't support large UID/GID values, so set them to 0 to
avoid a hard error from at least GNU tar
* OpenBSD tar needs -F ustar, and it appears to warn but carry
on with "nobody" if a UID is too large
* -f /dev/null is a more portable way to throw away the output, since
the default destination might be a tape device depending on build
options that a distribution might change
* Windows ships BSD tar but lacks /dev/null, so ask perl for its name
Based on their manuals, the other two implementations the tests are
likely to encounter in the wild don't seem to need any special handling:
* Solaris/illumos tar uses ustar and replaces large UIDs with 60001
* AIX tar uses ustar (unless --format=pax) and truncates large UIDs
Backpatch-through: 18
Co-authored-by: Thomas Munro <thomas.munro@gmail.com>
Co-authored-by: Sami Imseih <samimseih@gmail.com> (large UIDs)
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us> (earlier version)
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com> (OpenBSD)
Reviewed-by: Andrew Dunstan <andrew@dunslane.net> (Windows)
Discussion: https://postgr.es/m/3676229.1775170250%40sss.pgh.pa.us
Discussion: https://postgr.es/m/CAA5RZ0tt89MgNi4-0F4onH%2B-TFSsysFjMM-tBc6aXbuQv5xBXw%40mail.gmail.com
It's not very useful to specify a non-standard directory size. The
HASH_DIRSIZE option was only used for shared memory hash tables, and
those always used hash_select_dirsize() to choose the size, which in
turn just uses the default algorithm anyway. That assumption was
ingrained in hash_estimate_size(), too.
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://www.postgresql.org/message-id/01ab1d41-3eda-4705-8bbd-af898f5007f1@iki.fi
Previously, the shared header (HASHHDR) and the directory were
allocated by the caller, and passed to hash_create(), while the actual
elements were allocated separately with ShmemAlloc(). After this
commit, all the memory needed by the header, the directory, and all
the elements is allocated using a single ShmemInitStruct() call, and
the different parts are carved out of that allocation. This way the
ShmemIndex entries (and thus pg_shmem_allocations) reflect the size of
the whole hash table, rather than just the directories.
Commit f5930f9a98 attempted this earlier, but it had to be reverted.
The new strategy is to let dynahash.c perform all the allocations with
the alloc function, but have the alloc function carve out the parts
from the one larger allocation. The shared header and the directory
are now also allocated with alloc calls, instead of passing the area
for those directly from the caller.
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://www.postgresql.org/message-id/01ab1d41-3eda-4705-8bbd-af898f5007f1@iki.fi
Set HASH_FIXED_SIZE on all shared memory hash tables, to prevent them
from growing after the initial allocation. It was always weirdly
indeterministic that if one hash table used up all the unused shared
memory, you could not use that space for other things anymore until
restart. We just got rid of that behavior for the LOCK and PROCLOCK
tables, but it's similarly weird for all other hash tables.
Increase SHMEM_INDEX_SIZE because we were already above the max size,
on that one, and it's now a hard limit.
Some callers of ShmemInitHash() still pass HASH_FIXED_SIZE, but that's
now unnecessary. They should perhaps now be removed, but it doesn't do
any harm either to pass it.
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://www.postgresql.org/message-id/01ab1d41-3eda-4705-8bbd-af898f5007f1@iki.fi
Replace the separate init and max size options with a single size
option. We didn't make much use of the feature, all callers except the
ones in wait_event.c already used the same size for both, and the hash
tables in wait_event.c are small so there's little harm in just
allocating them to the max size.
The only reason why you might want to not reserve the max size upfront
is to make the memory available for other hash tables to grow beyond
their max size. Letting hash tables grow much beyond their max size is
bad for performance, however, because we cannot resize the directory,
and we never had very much "wiggle room" to grow to anyway so you
couldn't really rely on it. We recently marked the LOCK and PROCLOCK
tables with HAS_FIXED_SIZE, so there's nothing left in core that would
benefit from more unallocated shared memory.
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://www.postgresql.org/message-id/01ab1d41-3eda-4705-8bbd-af898f5007f1@iki.fi
At the moment, the only way for a validator module to report error
details on failure is to log them separately before returning from
validate_cb. Independently of that problem, the ereport() calls that we
make during validation failure partially duplicate some of the work of
auth_failed().
The end result is overly verbose and confusing for readers of the logs:
[768233] LOG: [my_validator] bad signature in bearer token
[768233] LOG: OAuth bearer authentication failed for user "jacob"
[768233] DETAIL: Validator failed to authorize the provided token.
[768233] FATAL: OAuth bearer authentication failed for user "jacob"
[768233] DETAIL: Connection matched file ".../pg_hba.conf" line ...
Solve both problems by making use of the existing logdetail pointer
that's provided by ClientAuthentication. Validator modules may set
ValidatorModuleResult->error_detail to override our default generic
message.
The end result looks something like
[242284] FATAL: OAuth bearer authentication failed for user "jacob"
[242284] DETAIL: [my_validator] bad signature in bearer token
Connection matched file ".../pg_hba.conf" line ...
Reported-by: Álvaro Herrera <alvherre@kurilemu.de>
Reported-by: Zsolt Parragi <zsolt.parragi@percona.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Zsolt Parragi <zsolt.parragi@percona.com>
Discussion: https://postgr.es/m/202601241015.y5uvxd7oxnfs%40alvherre.pgsql
The test for re-running checksum enabling was only checking for the
data checksum state to transition to 'on', but didn't account for
the launcher process having had time to exit, thus getting an error
instead of the expected no-op. Adding a pg_stat_activity check for
the launcher exiting resolves the error, verified by inducing delay
in the launcher.
Also wrap a variable only used in injection point tests within the
correct USE macros to avoid warning for an unused variable.
All per the buildfarm.
Author: Daniel Gustafsson <daniel@yesql.se>
Reported-by: Buildfarm
Discussion: https://postgr.es/m/1CB288C9-564B-4664-B096-C2F4377D17AB@yesql.se
Presently, this function only computes component scores when the
corresponding threshold is reached. A follow-up commit will add a
view that shows tables' autovacuum scores, and we anticipate that
users will want to use this view to discover tables that are
nearing autovacuum eligibility. This commit teaches this function
to always compute autovacuum scores, even when a threshold has not
been reached or autovacuum is disabled.
The restructuring in this commit revealed an interesting edge case.
If the table needs vacuuming for wraparound prevention and
autovacuum is disabled for it, we might still choose to analyze it.
It's not clear if this is intentional, but it has been this way for
nearly 20 years, so it seems best to avoid changing it without
further discussion.
Author: Sami Imseih <samimseih@gmail.com>
Reviewed-by: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Discussion: https://postgr.es/m/CAA5RZ0s4xjMrB-VAnLccC7kY8d0-4806-Lsac-czJsdA1LXtAw%40mail.gmail.com
This allows data checksums to be enabled, or disabled, in a running
cluster without restricting access to the cluster during processing.
Data checksums could prior to this only be enabled during initdb or
when the cluster is offline using the pg_checksums app. This commit
introduce functionality to enable, or disable, data checksums while
the cluster is running regardless of how it was initialized.
A background worker launcher process is responsible for launching a
dynamic per-database background worker which will mark all buffers
dirty for all relation with storage in order for them to have data
checksums calculated on write. Once all relations in all databases
have been processed, the data_checksums state will be set to on and
the cluster will at that point be identical to one which had data
checksums enabled during initialization or via offline processing.
When data checksums are being enabled, concurrent I/O operations
from backends other than the data checksums worker will write the
checksums but not verify them on reading. Only when all backends
have absorbed the procsignalbarrier for setting data_checksums to
on will they also start verifying checksums on reading. The same
process is repeated during disabling; all backends write checksums
but do not verify them until the barrier for setting the state to
off has been absorbed by all. This in-progress state is used to
ensure there are no false negatives (or positives) due to reading
a checksum which is not in sync with the page.
A new testmodule, test_checksums, is introduced with an extensive
set of tests covering both online and offline data checksum mode
changes. The tests which run concurrent pgbdench during online
processing are gated behind the PG_TEST_EXTRA flag due to being
very expensive to run. Two levels of PG_TEST_EXTRA flags exist
to turn on a subset of the expensive tests, or the full suite of
multiple runs.
This work is based on an earlier version of this patch which was
reviewed by among others Heikki Linnakangas, Robert Haas, Andres
Freund, Tomas Vondra, Michael Banck and Andrey Borodin. During
the work on this new version, Tomas Vondra has given invaluable
assistance with not only coding and reviewing but very in-depth
testing.
Author: Daniel Gustafsson <daniel@yesql.se>
Author: Magnus Hagander <magnus@hagander.net>
Co-authored-by: Tomas Vondra <tomas@vondra.me>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/CABUevExz9hUUOLnJVr2kpw9Cx=o4MCr1SVKwbupzuxP7ckNutA@mail.gmail.com
Discussion: https://postgr.es/m/20181030051643.elbxjww5jjgnjaxg@alap3.anarazel.de
Discussion: https://postgr.es/m/CABUevEwE3urLtwxxqdgd5O2oQz9J717ZzMbh+ziCSa5YLLU_BA@mail.gmail.com
This commit adds an early return to this function, allowing us to
remove a level of indentation on a decent chunk of code. This is
preparatory work for follow-up commits that will add a new system
view to show tables' autovacuum scores.
Reviewed-by: Sami Imseih <samimseih@gmail.com>
Discussion: https://postgr.es/m/CAA5RZ0s4xjMrB-VAnLccC7kY8d0-4806-Lsac-czJsdA1LXtAw%40mail.gmail.com
The previous commits reduced the amount of memory available for locks
by eliminating the "safety margins" and by settling the split between
LOCK and PROCLOCK tables at startup. The allocation is now more
deterministic, but it also means that you often hit one of the limits
sooner than before. To compensate for that, bump up
max_locks_per_transactions from 64 to 128. With that there is a little
more space in the both hash tables than what was the effective maximum
size for either table before the previous commits.
This only changes the default, so if you had changed
max_locks_per_transactions in postgresql.conf, you will still have
fewer locks available than before for the same setting value. This
should be noted in the release notes. A good rule of thumb is that if
you double max_locks_per_transactions, you should be able to get as
many locks as before.
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://www.postgresql.org/message-id/e07be2ba-856b-4ff5-8313-8b58b6b4e4d0@iki.fi
This prevents the LOCK table from "stealing" space that was originally
calculated for the PROLOCK table, and vice versa. That was weirdly
indeterministic so that if you e.g. took a lot of locks consuming all
the available shared memory for the LOCK table, subsequent
transactions that needed the more space for the PROCLOCK table would
fail, but if you restarted the system then the space would be
available for PROCLOCK again. Better to be strict and predictable,
even though that means that in many cases you can acquire far fewer
locks than before.
This also prevents the lock hash tables from using up the
general-purpose 100 kB reserve we set aside for "stuff that's too
small to bother estimating" in CalculateShmemSize(). We are pretty
good at accounting for everything nowadays, so we could probably make
that reservation smaller, but I'll leave that for another commit.
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://www.postgresql.org/message-id/e07be2ba-856b-4ff5-8313-8b58b6b4e4d0@iki.fi
As the comment says, the hash table sizes are just estimates, but that
doesn't mean we need a "safety margin" here. hash_estimate_size()
estimates the needed size in bytes pretty accurately for the given
number of elements, so if we wanted room for more elements in the
table, we should just use larger max_table_size in the
hash_estimate_size() call.
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://www.postgresql.org/message-id/e07be2ba-856b-4ff5-8313-8b58b6b4e4d0@iki.fi
The 10% safety margin was copy-pasted from lock.c when the predicate
locking code was originally added. However, we later (commit
7c797e7194) added the HASH_FIXED_SIZE flag to the hash tables, which
means that they cannot actually use the safety margin that we're
calculating for them.
The extra memory was mainly used by the main lock manager, which is
the only shmem hash table of non-trivial size that does not use the
HASH_FIXED_SIZE flag. If we wanted to have more space for the lock
manager, we should reserve it directly in lock.c. After this commit,
the lock manager will just have less memory available than before.
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://www.postgresql.org/message-id/e07be2ba-856b-4ff5-8313-8b58b6b4e4d0@iki.fi
Instead of probing the PK index on each trigger invocation, buffer
FK rows in a new per-constraint cache entry (RI_FastPathEntry) and
flush them as a batch.
On each trigger invocation, the new ri_FastPathBatchAdd() buffers
the FK row in RI_FastPathEntry. When the buffer fills (64 rows)
or the trigger-firing cycle ends, the new ri_FastPathBatchFlush()
probes the index for all buffered rows, sharing a single
CommandCounterIncrement, snapshot, permission check, and security
context switch across the batch, rather than repeating each per row
as the SPI path does. Per-flush CCI is safe because all AFTER
triggers for the buffered rows have already fired by flush time.
For single-column foreign keys, the new ri_FastPathFlushArray()
builds an ArrayType from the buffered FK values (casting to the
PK-side type if needed) and constructs a scan key with the
SK_SEARCHARRAY flag. The index AM sorts and deduplicates the array
internally, then walks matching leaf pages in one ordered traversal
instead of descending from the root once per row. A matched[] bitmap
tracks which batch items were satisfied; the first unmatched item is
reported as a violation. Multi-column foreign keys fall back to
per-row probing via the new ri_FastPathFlushLoop().
The fast path introduced in the previous commit (2da86c1ef9) yields
~1.8x speedup. This commit adds ~1.6x on top of that, for a combined
~2.9x speedup over the unpatched code (int PK / int FK, 1M rows, PK
table and index cached in memory).
FK tuples are materialized via ExecCopySlotHeapTuple() into a new
purpose-specific memory context (flush_cxt), child of
TopTransactionContext, which is also used for per-flush transient
work: cast results, the search array, and index scan allocations.
It is reset after each flush and deleted in teardown.
The PK relation, index, tuple slots, and fast-path metadata are
cached in RI_FastPathEntry across trigger invocations within a
trigger-firing batch, avoiding repeated open/close overhead. The
snapshot and IndexScanDesc are taken fresh per flush. The entry is
not subject to cache invalidation: cached relations are held with
locks for the transaction duration, and the entry's lifetime is
bounded by the trigger-firing cycle.
Lifecycle management for RI_FastPathEntry relies on three new
mechanisms:
- AfterTriggerBatchCallback: A new general-purpose callback
mechanism in trigger.c. Callbacks registered via
RegisterAfterTriggerBatchCallback() fire at the end of each
trigger-firing batch (AfterTriggerEndQuery for immediate
constraints, AfterTriggerFireDeferred at COMMIT, and
AfterTriggerSetState for SET CONSTRAINTS IMMEDIATE). The RI
code registers ri_FastPathEndBatch as a batch callback.
- Batch callbacks only fire at the outermost query level
(checked inside FireAfterTriggerBatchCallbacks), so nested
queries from SPI inside other AFTER triggers do not tear down
the cache mid-batch.
- XactCallback: ri_FastPathXactCallback NULLs the static cache
pointer at transaction end, handling the abort path where the
batch callback never fired.
- SubXactCallback: ri_FastPathSubXactCallback NULLs the static
cache pointer on subtransaction abort, preventing the batch
callback from accessing already-released resources.
- AfterTriggerBatchIsActive(): A new exported accessor that
returns true when afterTriggers.query_depth >= 0. During
ALTER TABLE ... ADD FOREIGN KEY validation, RI triggers are
called directly outside the after-trigger framework, so batch
callbacks would never fire. The fast-path code uses this to
fall back to the non-cached per-invocation path in that
context.
ri_FastPathEndBatch() flushes any partial batch before tearing
down cached resources. Since the FK relation may already be
closed by flush time (e.g. for deferred constraints at COMMIT),
it reopens the relation using entry->fk_relid if needed.
The existing ALTER TABLE validation path bypasses batching and
continues to call ri_FastPathCheck() directly per row, because
RI triggers are called outside the after-trigger framework there
and batch callbacks would never fire to flush the buffer.
Suggested-by: David Rowley <dgrowleyml@gmail.com>
Author: Amit Langote <amitlangote09@gmail.com>
Co-authored-by: Junwang Zhao <zhjwpku@gmail.com>
Reviewed-by: Haibo Yan <tristan.yim@gmail.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Tested-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/CA+HiwqF4C0ws3cO+z5cLkPuvwnAwkSp7sfvgGj3yQ=Li6KNMqA@mail.gmail.com
LLVM 22 has the fix that we copied into our tree in commit 9044fc1d and
a new function to reach it[1][2], so we only need to use our copy for
Aarch64 + LLVM < 22. The only change to the final version that our copy
didn't get is a new LLVM_ABI macro, but that isn't appropriate for us.
Our copy is hopefully now frozen and would only need maintenance if bugs
are found in the upstream code.
Non-Aarch64 systems now also use the new API with LLVM 22. It allocates
all sections with one contiguous mmap() instead of one per
section. We could have done that earlier, but commit 9044fc1d wanted to
limit the blast radius to the affected systems. We might as well
benefit from that small improvement everywhere now that it is available
out of the box.
We can't delete our copy until LLVM 22 is our minimum supported version,
or we switch to the newer JITLink API for at least Aarch64.
[1] https://github.com/llvm/llvm-project/pull/71968
[2] https://github.com/llvm/llvm-project/pull/174307
Backpatch-through: 14
Discussion: https://postgr.es/m/CA%2BhUKGJTumad75o8Zao-LFseEbt%3DenbUFCM7LZVV%3Dc8yg2i7dg%40mail.gmail.com
Buildfarm testing shows that OpenSUSE (and perhaps related platforms?)
configures GNU tar in such a way that it'll archive sparse WAL files
by default, thus triggering the pax-extension detection code added by
bc30c704a. Thus, we need something similar to 852de579a but for
GNU tar's option set. "--format=ustar" seems to do the trick.
Moreover, the buildfarm shows that pg_verifybackup's 003_corruption.pl
test script is also triggering creation of pax-format tar files on
that platform. We had not noticed because those test cases all fail
(intentionally) before getting to the point of trying to verify WAL
data.
Since that means two TAP scripts need this option-selection logic, and
plausibly more will do so in future, factor it out into a subroutine
in Test::Utils. We also need to back-patch the 003_corruption.pl fix
into v18, where it's also failing.
While at it, clean up some places where guards for $tar being empty
or undefined were incomplete or even outright backwards. Presumably,
we missed noticing because the set of machines that run TAP tests
and don't have tar installed is empty. But if we're going to try
to handle that scenario, we should do it correctly.
Reported-by: Tomas Vondra <tomas@vondra.me>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/02770bea-b3f3-4015-8a43-443ae345379c@vondra.me
Backpatch-through: 18
Add the following jsonpath methods:
* l/r/btrim()
* lower(), upper()
* initcap()
* replace()
* split_part()
Each simply dispatches to the standard string processing functions.
These depend on the locale, but since it's set at `initdb`, they can be
considered immutable and therefore allowed in any jsonpath expression.
Author: Florents Tselai <florents.tselai@gmail.com>
Co-authored-by: David E. Wheeler <david@justatheory.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Discussion: https://postgr.es/m/CA+v5N40sJF39m0v7h=QN86zGp0CUf9F1WKasnZy9nNVj_VhCZQ@mail.gmail.com
This is just cleanup in the jsonpath grammar.
Rename the `csv_` tokens to `int_`, because they represent signed or
unsigned integers, as follows:
* `csv_elem` => `int_elem`
* `csv_list` => `int_list`
* `opt_csv_list` => `opt_int_list`
Rename the `datetime_precision` tokens to `uint_arg`, as they represent
unsigned integers and will be useful for other methods in the future, as
follows:
* `datetime_precision` => `uint_elem`
* `opt_datetime_precision` => `opt_uint_arg`
Rename the `datetime_template` tokens to `str_arg`, as they represent
strings and will be useful for other methods in the future, as follows:
* `datetime_template` => `str_elem`
* `opt_datetime_template` => `opt_str_arg`
Author: David E. Wheeler <david@justatheory.com>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Discussion: https://postgr.es/m/CA+v5N40sJF39m0v7h=QN86zGp0CUf9F1WKasnZy9nNVj_VhCZQ@mail.gmail.com
When a tablesync worker checks whether a specific table is published,
it previously issued a query to the publisher calling
pg_get_publication_tables() and filtering the result by relid via a
WHERE clause. Because the function itself was fully evaluated before
the filter was applied, this forced the publisher to enumerate all
tables in the publication. For publications covering a large number of
tables, this resulted in expensive catalog scans and unnecessary CPU
overhead on the publisher.
This commit adds a new overloaded form of pg_get_publication_tables()
that accepts an array of publication names and a target table
OID. Instead of enumerating all published tables, it evaluates
membership for the specified relation via syscache lookups, using the
new is_table_publishable_in_publication() helper. This helper
correctly accounts for publish_via_partition_root, ALL TABLES with
EXCEPT clauses, schema publications, and partition inheritance, while
avoiding the overhead of building the complete published table list.
The existing VARIADIC array form of pg_get_publication_tables() is
preserved for backward compatibility. Tablesync workers use the new
two-argument form when connected to a publisher running PostgreSQL 19
or later.
Bump catalog version.
Reported-by: Marcos Pegoraro <marcos@f10.com.br>
Reviewed-by: Zhijie Hou <houzj.fnst@fujitsu.com>
Reviewed-by: Matheus Alcantara <matheusssilv97@gmail.com>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Reviewed-by: Peter Smith <smithpb2250@gmail.com>
Reviewed-by: Hayato Kuroda <kuroda.hayato@fujitsu.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Haoyan Wang <wanghaoyan20@163.com>
Discussion: https://postgr.es/m/CAB-JLwbBFNuASyEnZWP0Tck9uNkthBZqi6WoXNevUT6+mV8XmA@mail.gmail.com
Previously, there was essentially no verification in this code that
the input is a tar file at all, let alone that it fits into the
subset of valid tar files that we can handle. This was exposed by
the discovery that we couldn't handle files that FreeBSD's tar
makes, because it's fairly aggressive about converting sparse WAL
files into sparse tar entries. To fix:
* Bail out if we find a pax extension header. This covers the
sparse-file case, and also protects us against scenarios where
the pax header changes other file properties that we care about.
(Eventually we may extend the logic to actually handle such
headers, but that won't happen in time for v19.)
* Be more wary about tar file type codes in general: do not assume
that anything that's neither a directory nor a symlink must be a
regular file. Instead, we just ignore entries that are none of the
three supported types.
* Apply pg_dump's isValidTarHeader to verify that a purported
header block is actually in tar format. To make this possible,
move isValidTarHeader into src/port/tar.c, which is probably where
it should have been since that file was created.
I also took the opportunity to const-ify the arguments of
isValidTarHeader and tarChecksum, and to use symbols not hard-wired
constants inside tarChecksum.
Back-patch to v18 but not further. Although this code exists inside
pg_basebackup in older branches, it's not really exposed in that
usage to tar files that weren't generated by our own code, so it
doesn't seem worth back-porting these changes across 3c9056981
and f80b09bac. I did choose to include a back-patch of 5868372bb
into v18 though, to minimize cosmetic differences between these
two branches.
Author: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/3049460.1775067940@sss.pgh.pa.us>
Backpatch-through: 18
Interrupt handling functions (e.g., HandleCatchupInterrupt(),
HandleParallelApplyMessageInterrupt()) are called only by
procsignal_sigusr1_handler(), which already calls SetLatch()
for the current process at the end of its processing.
Therefore, these interrupt handling functions do not need to
call SetLatch() themselves.
However, previously, some of these functions redundantly
called SetLatch(). This commit removes those unnecessary
calls.
While duplicate SetLatch() calls are redundant, they are
harmless, so this change is not backpatched.
Author: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Reviewed-by: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Dilip Kumar <dilipbalaut@gmail.com>
Discussion: https://postgr.es/m/CALj2ACWd5apddj6Cd885WwJ6LquYu_G81C4GoR4xSoDV1x-FEA@mail.gmail.com
Previously we would only check for the availability of __cpuidex if
the related __get_cpuid_count was not available on a platform.
Future commits will need to access hypervisor information about
the TSC frequency of x86 CPUs. For that case __cpuidex is the only
viable option for accessing a high leaf (e.g. 0x40000000), since
__get_cpuid_count does not allow that.
__cpuidex is defined in cpuid.h for gcc/clang, but in intrin.h
for MSVC, so adjust tests to suite. We also need to cast the array
of unsigned ints to signed, since gcc (with -Wall) and clang emit
warnings otherwise.
Author: Lukas Fittl <lukas@fittl.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: John Naylor <john.naylor@postgresql.org>
Discussion: https://postgr.es/m/CAP53PkyooCeR8YV0BUD_xC7oTZESHz8OdA=tP7pBRHFVQ9xtKg@mail.gmail.com
Now that command_ok() captures and displays failure output, use it
instead of system() plus manual diff-dumping in these two tests. This
simplifies both scripts and produces consistent, truncated output on
failure.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Corey Huinker <corey.huinker@gmail.com>
Discussion: https://postgr.es/m/DFYFWM053WHS.10K8ZPJ605UFK@jeltef.nl
Replace die with croak throughout Cluster.pm and Utils.pm (except in
INIT blocks and signal handlers, where die is correct) so that error
messages report the test script's line number rather than the helper
module's.
Add @CARP_NOT in Utils.pm listing PostgreSQL::Test::Cluster, so that
when a Utils function is called through a Cluster.pm wrapper, croak
skips both packages and reports the actual test-script caller.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Corey Huinker <corey.huinker@gmail.com>
Reviewed-by: Zsolt Parragi <zsolt.parragi@percona.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/DFYFWM053WHS.10K8ZPJ605UFK@jeltef.nl
Install a $SIG{__DIE__} handler in the INIT block of Utils.pm that emits
the die message as a TAP diagnostic. Previously, an unexpected die
(e.g. from safe_psql) produced only "no plan was declared" with no
indication of the actual error. The handler also calls done_testing()
to suppress that confusing message.
Dies during compilation ($^S undefined) and inside eval ($^S == 1) are
left alone.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Corey Huinker <corey.huinker@gmail.com>
Reviewed-by: Zsolt Parragi <zsolt.parragi@percona.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/DFYFWM053WHS.10K8ZPJ605UFK@jeltef.nl
Discussion: https://postgr.es/m/20220222181924.eehi7o4pmneeb4hm%40alap3.anarazel.de
Capture stdout and stderr from command_ok() and command_fails() and emit
them as TAP diagnostics on failure. Output is truncated to the first
and last 30 lines per channel to avoid flooding.
A new helper _diag_command_output() is introduced in Utils.pm so
both functions share the same truncation and formatting logic.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Corey Huinker <corey.huinker@gmail.com>
Reviewed-by: Zsolt Parragi <zsolt.parragi@percona.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/DFYFWM053WHS.10K8ZPJ605UFK@jeltef.nl
When pg_regress fails it is often tedious to find the actual diffs,
especially in CI where you must navigate a file browser. Emit the first
80 lines of the combined regression.diffs as TAP diagnostics so the
failure reason is visible directly in the test output.
The line limit is across all failing tests in a single pg_regress run to
avoid flooding when a crash causes every subsequent test to fail.
New DIAG_DETAIL / DIAG_END tap output types are added, mirroring the
existing NOTE_DETAIL / NOTE_END pair, so that long diff lines can be
emitted without spurious '#' prefixes on continuation lines.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Corey Huinker <corey.huinker@gmail.com>
Reviewed-by: Zsolt Parragi <zsolt.parragi@percona.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/DFYFWM053WHS.10K8ZPJ605UFK@jeltef.nl
While JIT can speed up large analytical queries, it can also cause
serious performance issues on otherwise very fast queries. Compiling
and optimizing the expressions may be so expensive, it completely
outweighs the JIT benefits for shorter queries.
Ideally, we'd address this in the cost model, but the part deciding
whether to enable JIT for a query is rather simple, partially because we
don't have any reliable estimates of how expensive the LLVM compilation
and optimization is.
Sometimes seemingly unrelated changes (for example a couple additional
INSERTs into a table) increase the cost just enough to enable JIT,
resulting in a performance cliff.
Because of these risks, most large-scale deployments already disable JIT
by default. Notably, this includes all hyperscalers.
This commit changes our default to align with that established practice.
If we improve the JIT (be it better costing or cheaper execution), we
can consider enabling it by default again.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Discussion: https://postgr.es/m/DG1VZJEX1AQH.2EH4OKGRUDB71@jeltef.nl
