In addition to the added functions, the pg_buffercache_evict() function now
shows whether the buffer was flushed.
pg_buffercache_evict_relation(): Evicts all shared buffers in a
relation at once.
pg_buffercache_evict_all(): Evicts all shared buffers at once.
Both functions provide mechanism to evict multiple shared buffers at
once. They are designed to address the inefficiency of repeatedly calling
pg_buffercache_evict() for each individual buffer, which can be time-consuming
when dealing with large shared buffer pools. (e.g., ~477ms vs. ~2576ms for
16GB of fully populated shared buffers).
These functions are intended for developer testing and debugging
purposes and are available to superusers only.
Minimal tests for the new functions are included. Also, there was no test for
pg_buffercache_evict(), test for this added too.
No new extension version is needed, as it was already increased this release
by ba2a3c2302f.
Author: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Aidar Imamov <a.imamov@postgrespro.ru>
Reviewed-by: Joseph Koshakow <koshy44@gmail.com>
Discussion: https://postgr.es/m/CAN55FZ0h_YoSqqutxV6DES1RW8ig6wcA8CR9rJk358YRMxZFmw%40mail.gmail.com
When planning queries to partitioned tables, we clone all
EquivalenceMembers belonging to the partitioned table into em_is_child
EquivalenceMembers for each non-pruned partition. For partitioned tables
with large numbers of partitions, this meant the ec_members list could
become large and code searching that list would become slow. Effectively,
the more partitions which were present, the more searches needed to be
performed for operations such as find_ec_member_matching_expr() during
create_plan() and the more partitions present, the longer these searches
would take, i.e., a quadratic slowdown.
To fix this, here we adjust how we store EquivalenceMembers for
em_is_child members. Instead of storing these directly in ec_members,
these are now stored in a new array of Lists in the EquivalenceClass,
which is indexed by the relid. When we want to find EquivalenceMembers
belonging to a certain child relation, we can narrow the search to the
array element for that relation.
To make EquivalenceMember lookup easier and to reduce the amount of code
change, this commit provides a pair of functions to allow iteration over
the EquivalenceMembers of an EC which also handles finding the child
members, if required. Callers that never need to look at child members
can remain using the foreach loop over ec_members, which will now often
be faster due to only parent-level members being stored there.
The actual performance increases here are highly dependent on the number
of partitions and the query being planned. Performance increases can be
visible with as few as 8 partitions, but the speedup is marginal for
such low numbers of partitions. The speedups become much more visible
with a few dozen to hundreds of partitions. With some tested queries
using 56 partitions, the planner was around 3x faster than before. For
use cases with thousands of partitions, these are likely to become
significantly faster. Some testing has shown planner speedups of 60x or
more with 8192 partitions.
Author: Yuya Watari <watari.yuya@gmail.com>
Co-authored-by: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Andrey Lepikhov <a.lepikhov@postgrespro.ru>
Reviewed-by: Alena Rybakina <lena.ribackina@yandex.ru>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Amit Langote <amitlangote09@gmail.com>
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Tested-by: Thom Brown <thom@linux.com>
Tested-by: newtglobal postgresql_contributors <postgresql_contributors@newtglobalcorp.com>
Discussion: https://postgr.es/m/CAJ2pMkZNCgoUKSE%2B_5LthD%2BKbXKvq6h2hQN8Esxpxd%2Bcxmgomg%40mail.gmail.com
Introduce new pg_shmem_alloctions_numa view with information about how
shared memory is distributed across NUMA nodes. For each shared memory
segment, the view returns one row for each NUMA node backing it, with
the total amount of memory allocated from that node.
The view may be relatively expensive, especially when executed for the
first time in a backend, as it has to touch all memory pages to get
reliable information about the NUMA node. This may also force allocation
of the shared memory.
Unlike pg_shmem_allocations, the view does not show anonymous shared
memory allocations. It also does not show memory allocated using the
dynamic shared memory infrastructure.
Author: Jakub Wartak <jakub.wartak@enterprisedb.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/CAKZiRmxh6KWo0aqRqvmcoaX2jUxZYb4kGp3N%3Dq1w%2BDiH-696Xw%40mail.gmail.com
Add basic NUMA awareness routines, using a minimal src/port/pg_numa.c
portability wrapper and an optional build dependency, enabled by
--with-libnuma configure option. For now this is Linux-only, other
platforms may be supported later.
A built-in SQL function pg_numa_available() allows checking NUMA
support, i.e. that the server was built/linked with the NUMA library.
The main function introduced is pg_numa_query_pages(), which allows
determining the NUMA node for individual memory pages. Internally the
function uses move_pages(2) syscall, as it allows batching, and is more
efficient than get_mempolicy(2).
Author: Jakub Wartak <jakub.wartak@enterprisedb.com>
Co-authored-by: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/CAKZiRmxh6KWo0aqRqvmcoaX2jUxZYb4kGp3N%3Dq1w%2BDiH-696Xw%40mail.gmail.com
In some edge cases valgrind flags issues with the memory referenced by
IOs. All of the cases addressed in this change are false positives.
Most of the false positives are caused by UnpinBuffer[NoOwner] marking buffer
data as inaccessible. This happens even though the AIO subsystem still holds a
pin. That's good, there shouldn't be accesses to the buffer outside of AIO
related code until it is pinned by "user" code again. But it requires some
explicit work - if the buffer is not pinned by the current backend, we need to
explicitly mark the buffer data accessible/inaccessible while executing
completion callbacks.
That however causes a cascading issue in IO workers: After the completion
callbacks for a buffer is executed, the page is marked as inaccessible. If
subsequently the same worker is executing IO targeting the same buffer, we
would get an error, as the memory is still marked inaccessible. To avoid that,
we need to explicitly mark the memory as accessible in IO workers.
Another issue is that IO executed in workers or via io_uring will not mark
memory as DEFINED. In the case of workers that is because valgrind does not
track memory definedness across processes. For io_uring that is because
valgrind does not understand io_uring, and therefore its IOs never mark memory
as defined, whether the completions are processed in the defining process or
in another context. It's not entirely clear how to best solve that. The
current user of AIO is not affected, as it explicitly marks buffers as DEFINED
& NOACCESS anyway. Defer solving this issue until we have a user with
different needs.
Per buildfarm animal skink.
Reviewed-by: Noah Misch <noah@leadboat.com>
Co-authored-by: Noah Misch <noah@leadboat.com>
Discussion: https://postgr.es/m/3pd4322mogfmdd5nln3zphdwhtmq3rzdldqjwb2sfqzcgs22lf@ok2gletdaoe6
The list of transform objects that a function should use is specified
in CREATE FUNCTION's TRANSFORM clause, and then represented indirectly
in pg_proc.protrftypes. However, ProcedureCreate completely ignored
that for purposes of constructing pg_depend entries, and instead made
the function depend on any transforms that exist for its parameter or
return data types. This is bad in both directions: the function could
be made dependent on a transform it does not actually use, or it
could try to use a transform that's since been dropped. (The latter
scenario would require use of a transform that's not for any of the
parameter or return types, but that seems legit for cases where the
function performs SQL operations internally.)
To fix, pass in the list of transform objects that CreateFunction
identified, and build pg_depend entries from that not from the
parameter/return types. This results in changes in the expected
test outputs in contrib/bool_plperl, which I guess are due to
different ordering of pg_depend entries -- that test case is
surely not exercising either of the problem scenarios.
This fix is not back-patchable as-is: changing the signature of
ProcedureCreate seems too risky in stable branches. We could
do something like making ProcedureCreate a wrapper around
ProcedureCreateExt or so. However, I'm more inclined to do
nothing in the back branches. We had no field complaints up to
now, so the hazards don't seem to be a big issue in practice.
And we couldn't do anything about existing pg_depend entries,
so a back-patched fix would result in a mishmash of dependencies
created according to different rules. That cure could be worse
than the disease, perhaps.
I bumped catversion just to lay down a marker that the expected
contents of pg_depend are a bit different than before.
Reported-by: Chapman Flack <jcflack@acm.org>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/3112950.1743984111@sss.pgh.pa.us
This allows them to be added without scanning the table, and validating
them afterwards without holding access exclusive lock on the table after
any violating rows have been deleted or fixed.
Doing ALTER TABLE ... SET NOT NULL for a column that has an invalid
not-null constraint validates that constraint. ALTER TABLE .. VALIDATE
CONSTRAINT is also supported. There are various checks on whether an
invalid constraint is allowed in a child table when the parent table has
a valid constraint; this should match what we do for enforced/not
enforced constraints.
pg_attribute.attnotnull is now only an indicator for whether a not-null
constraint exists for the column; whether it's valid or invalid must be
queried in pg_constraint. Applications can continue to query
pg_attribute.attnotnull as before, but now it's possible that NULL rows
are present in the column even when that's set to true.
For backend internal purposes, we cache the nullability status in
CompactAttribute->attnullability that each tuple descriptor carries
(replacing CompactAttribute.attnotnull, which was a mirror of
Form_pg_attribute.attnotnull). During the initial tuple descriptor
creation, based on the pg_attribute scan, we set this to UNRESTRICTED if
pg_attribute.attnotnull is false, or to UNKNOWN if it's true; then we
update the latter to VALID or INVALID depending on the pg_constraint
scan. This flag is also copied when tupledescs are copied.
Comparing tuple descs for equality must also compare the
CompactAttribute.attnullability flag and return false in case of a
mismatch.
pg_dump deals with these constraints by storing the OIDs of invalid
not-null constraints in a separate array, and running a query to obtain
their properties. The regular table creation SQL omits them entirely.
They are then dealt with in the same way as "separate" CHECK
constraints, and dumped after the data has been loaded. Because no
additional pg_dump infrastructure was required, we don't bump its
version number.
I decided not to bump catversion either, because the old catalog state
works perfectly in the new world. (Trying to run with new catalog state
and the old server version would likely run into issues, however.)
System catalogs do not support invalid not-null constraints (because
commit 14e87ffa5c54 didn't allow them to have pg_constraint rows
anyway.)
Author: Rushabh Lathia <rushabh.lathia@gmail.com>
Author: Jian He <jian.universality@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Tested-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Discussion: https://postgr.es/m/CAGPqQf0KitkNack4F5CFkFi-9Dqvp29Ro=EpcWt=4_hs-Rt+bQ@mail.gmail.com
This escape shows the numeric server IP address that the client
has connected to. Unix-socket connections will show "[local]".
Non-client processes (e.g. background processes) will show "[none]".
We expect that this option will be of interest to only a fairly
small number of users. Therefore the implementation is optimized
for the case where it's not used (that is, we don't do the string
conversion until we have to), and we've not added the field to
csvlog or jsonlog formats.
Author: Greg Sabino Mullane <htamfids@gmail.com>
Reviewed-by: Cary Huang <cary.huang@highgo.ca>
Reviewed-by: David Steele <david@pgmasters.net>
Reviewed-by: Jim Jones <jim.jones@uni-muenster.de>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CAKAnmmK-U+UicE-qbNU23K--Q5XTLdM6bj+gbkZBZkjyjrd3Ow@mail.gmail.com
There were several places in ordering-related planning where a
requirement for btree was hardcoded but an amcanorder index could
suffice. This fixes that. We just need to do the necessary mapping
between strategy numbers and compare types and adjust some related
APIs so that this works independent of btree strategy numbers. For
instance, non-btree amcanorder indexes can now be used to support
sorting and merge joins. Also, predtest.c works independent of btree
strategy numbers now.
To avoid performance regressions, some details on btree and other
built-in index types are still hardcoded as shortcuts, but other index
types now have access to the same features by providing the required
flags and callbacks.
Author: Mark Dilger <mark.dilger@enterprisedb.com>
Co-authored-by: Peter Eisentraut <peter@eisentraut.org>
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
The previous implementation of CRC32C on x86 relied on the native
CRC32 instruction from the SSE 4.2 extension, which operates on
up to 8 bytes at a time. We can get a substantial speedup by using
carryless multiplication on SIMD registers, processing 64 bytes per
loop iteration. Shorter inputs fall back to ordinary CRC instructions.
On Intel Tiger Lake hardware (2020), CRC is now 50% faster for inputs
between 64 and 112 bytes, and 3x faster for 256 bytes.
The VPCLMULQDQ instruction on 512-bit registers has been available
on Intel hardware since 2019 and AMD since 2022. There is an older
variant for 128-bit registers, but at least on Zen 2 it performs worse
than normal CRC instructions for short inputs.
We must now do a runtime check, even for builds that target SSE
4.2. This doesn't matter in practice for WAL (arguably the most
critical case), because since commit e2809e3a1 the final computation
with the 20-byte WAL header is inlined and unrolled when targeting
that extension. Compared with two direct function calls, testing
showed equal or slightly faster performance in performing an indirect
function call on several dozen bytes followed by inlined instructions
on constant input of 20 bytes.
The MIT-licensed implementation was generated with the "generate"
program from
https://github.com/corsix/fast-crc32/
Based on: "Fast CRC Computation for Generic Polynomials Using PCLMULQDQ
Instruction" V. Gopal, E. Ozturk, et al., 2009
Co-authored-by: Raghuveer Devulapalli <raghuveer.devulapalli@intel.com>
Co-authored-by: Paul Amonson <paul.d.amonson@intel.com>
Reviewed-by: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de> (earlier version)
Reviewed-by: Matthew Sterrett <matthewsterrett2@gmail.com> (earlier version)
Tested-by: Raghuveer Devulapalli <raghuveer.devulapalli@intel.com>
Tested-by: David Rowley <<dgrowleyml@gmail.com>> (earlier version)
Discussion: https://postgr.es/m/BL1PR11MB530401FA7E9B1CA432CF9DC3DC192@BL1PR11MB5304.namprd11.prod.outlook.com
Discussion: https://postgr.es/m/PH8PR11MB82869FF741DFA4E9A029FF13FBF72@PH8PR11MB8286.namprd11.prod.outlook.com
Don't allow nbtree scans with skip arrays to end any primitive scan on
its first leaf page without giving some consideration to how many times
the scan's arrays advanced while changing at least one skip array
(though continue not caring about the number of array advancements that
only affected SAOP arrays, even during skip scans with SAOP arrays).
Now when a scan performs more than 3 such array advancements in the
course of reading a single leaf page, it is taken as a signal that the
next page is unlikely to be skippable. We'll therefore continue the
ongoing primitive index scan, at least until we can perform a recheck
against the next page's finaltup.
Testing has shown that this new heuristic occasionally makes all the
difference with skip scans that were expected to rely on the "passed
first page" heuristic added by commit 9a2e2a28. Without it, there is a
remaining risk that certain kinds of skip scans will never quite manage
to clear the initial hurdle of performing a primitive scan that lasts
beyond its first leaf page (or that such a skip scan will only clear
that initial hurdle when it has already wasted noticeably-many cycles
due to inefficient primitive scan scheduling).
Follow-up to commits 92fe23d9 and 9a2e2a28.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://postgr.es/m/CAH2-Wz=RVdG3zWytFWBsyW7fWH7zveFvTHed5JKEsuTT0RCO_A@mail.gmail.com
Postgres 17 commit e0b1ee17 added two complementary optimizations to
nbtree: the "prechecked" and "firstmatch" optimizations. _bt_readpage
was made to avoid needlessly evaluating keys that are guaranteed to be
satisfied by applying page-level context. "prechecked" did this for
keys required in the current scan direction, while "firstmatch" did it
for keys required in the opposite-to-scan direction only.
The "prechecked" design had a number of notable issues. It didn't
account for the fact that an = array scan key's sk_argument field might
need to advance at the point of the page precheck (it didn't check the
precheck tuple against the key's array, only the key's sk_argument,
which needlessly made it ineffective in cases involving stepping to a
page having advanced the scan's arrays using a truncated high key).
"prechecked" was also completely ineffective when only one scan key
wasn't guaranteed to be satisfied by every tuple (it didn't recognize
that it was still safe to avoid evaluating other, earlier keys).
The "firstmatch" optimization had similar limitations. It could only be
applied after _bt_readpage found its first matching tuple, regardless of
why any earlier tuples failed to satisfy the scan's index quals. This
allowed unsatisfied non-required scan keys to impede the optimization.
Replace both optimizations with a new optimization, without any of these
limitations: the "startikey" optimization. Affected _bt_readpage calls
generate a page-level key offset ("startikey"), that their _bt_checkkeys
calls can then start at. This is an offset to the first key that isn't
known to be satisfied by every tuple on the page.
Although this is independently useful work, its main goal is to avoid
performance regressions with index scans that use skip arrays, but still
never manage to skip over irrelevant leaf pages. We must avoid wasting
CPU cycles on overly granular skip array maintenance in these cases.
The new "startikey" optimization helps with this by selectively
disabling array maintenance for the duration of a _bt_readpage call.
This has no lasting consequences for the scan's array keys (they'll
still reliably track the scan's progress through the index's key space
whenever the scan is "between pages").
Skip scan adds skip arrays during preprocessing using simple, static
rules, and decides how best to navigate/apply the scan's skip arrays
dynamically, at runtime. The "startikey" optimization enables this
approach. As a result of all this, the planner doesn't need to generate
distinct, competing index paths (one path for skip scan, another for an
equivalent traditional full index scan). The overall effect is to make
scan runtime close to optimal, even when the planner works off an
incorrect cardinality estimate. Scans will also perform well given a
skipped column with data skew: individual groups of pages with many
distinct values (in respect of a skipped column) can be read about as
efficiently as before -- without the scan being forced to give up on
skipping over other groups of pages that are provably irrelevant.
Many scans that cannot possibly skip will still benefit from the use of
skip arrays, since they'll allow the "startikey" optimization to be as
effective as possible (by allowing preprocessing to mark all the scan's
keys as required). A scan that uses a skip array on "a" for a qual
"WHERE a BETWEEN 0 AND 1_000_000 AND b = 42" is often much faster now,
even when every tuple read by the scan has its own distinct "a" value.
However, there are still some remaining regressions, affecting certain
trickier cases.
Scans whose index quals have several range skip arrays, each on some
high cardinality column, can still be slower than they were before the
introduction of skip scan -- even with the new "startikey" optimization.
There are also known regressions affecting very selective index scans
that use a skip array. The underlying issue with such selective scans
is that they never get as far as reading a second leaf page, and so will
never get a chance to consider applying the "startikey" optimization.
In principle, all regressions could be avoided by teaching preprocessing
to not add skip arrays whenever they aren't expected to help, but it
seems best to err on the side of robust performance.
Follow-up to commit 92fe23d9, which added nbtree skip scan.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Reviewed-By: Masahiro Ikeda <ikedamsh@oss.nttdata.com>
Reviewed-By: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://postgr.es/m/CAH2-Wz=Y93jf5WjoOsN=xvqpMjRy-bxCE037bVFi-EasrpeUJA@mail.gmail.com
Discussion: https://postgr.es/m/CAH2-WznWDK45JfNPNvDxh6RQy-TaCwULaM5u5ALMXbjLBMcugQ@mail.gmail.com
Teach nbtree multi-column index scans to opportunistically skip over
irrelevant sections of the index given a query with no "=" conditions on
one or more prefix index columns. When nbtree is passed input scan keys
derived from a predicate "WHERE b = 5", new nbtree preprocessing steps
output "WHERE a = ANY(<every possible 'a' value>) AND b = 5" scan keys.
That is, preprocessing generates a "skip array" (and an output scan key)
for the omitted prefix column "a", which makes it safe to mark the scan
key on "b" as required to continue the scan. The scan is therefore able
to repeatedly reposition itself by applying both the "a" and "b" keys.
A skip array has "elements" that are generated procedurally and on
demand, but otherwise works just like a regular ScalarArrayOp array.
Preprocessing can freely add a skip array before or after any input
ScalarArrayOp arrays. Index scans with a skip array decide when and
where to reposition the scan using the same approach as any other scan
with array keys. This design builds on the design for array advancement
and primitive scan scheduling added to Postgres 17 by commit 5bf748b8.
Testing has shown that skip scans of an index with a low cardinality
skipped prefix column can be multiple orders of magnitude faster than an
equivalent full index scan (or sequential scan). In general, the
cardinality of the scan's skipped column(s) limits the number of leaf
pages that can be skipped over.
The core B-Tree operator classes on most discrete types generate their
array elements with the help of their own custom skip support routine.
This infrastructure gives nbtree a way to generate the next required
array element by incrementing (or decrementing) the current array value.
It can reduce the number of index descents in cases where the next
possible indexable value frequently turns out to be the next value
stored in the index. Opclasses that lack a skip support routine fall
back on having nbtree "increment" (or "decrement") a skip array's
current element by setting the NEXT (or PRIOR) scan key flag, without
directly changing the scan key's sk_argument. These sentinel values
behave just like any other value from an array -- though they can never
locate equal index tuples (they can only locate the next group of index
tuples containing the next set of non-sentinel values that the scan's
arrays need to advance to).
A skip array's range is constrained by "contradictory" inequality keys.
For example, a skip array on "x" will only generate the values 1 and 2
given a qual such as "WHERE x BETWEEN 1 AND 2 AND y = 66". Such a skip
array qual usually has near-identical performance characteristics to a
comparable SAOP qual "WHERE x = ANY('{1, 2}') AND y = 66". However,
improved performance isn't guaranteed. Much depends on physical index
characteristics.
B-Tree preprocessing is optimistic about skipping working out: it
applies static, generic rules when determining where to generate skip
arrays, which assumes that the runtime overhead of maintaining skip
arrays will pay for itself -- or lead to only a modest performance loss.
As things stand, these assumptions are much too optimistic: skip array
maintenance will lead to unacceptable regressions with unsympathetic
queries (queries whose scan can't skip over many irrelevant leaf pages).
An upcoming commit will address the problems in this area by enhancing
_bt_readpage's approach to saving cycles on scan key evaluation, making
it work in a way that directly considers the needs of = array keys
(particularly = skip array keys).
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Masahiro Ikeda <masahiro.ikeda@nttdata.com>
Reviewed-By: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Reviewed-By: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reviewed-By: Tomas Vondra <tomas@vondra.me>
Reviewed-By: Aleksander Alekseev <aleksander@timescale.com>
Reviewed-By: Alena Rybakina <a.rybakina@postgrespro.ru>
Discussion: https://postgr.es/m/CAH2-Wzmn1YsLzOGgjAQZdn1STSG_y8qP__vggTaPAYXJP+G4bw@mail.gmail.com
This commit implements the automatic conversion of 'x IN (VALUES ...)' into
ScalarArrayOpExpr. That simplifies the query tree, eliminating the appearance
of an unnecessary join.
Since VALUES describes a relational table, and the value of such a list is
a table row, the optimizer will likely face an underestimation problem due to
the inability to estimate cardinality through MCV statistics. The cardinality
evaluation mechanism can work with the array inclusion check operation.
If the array is small enough (< 100 elements), it will perform a statistical
evaluation element by element.
We perform the transformation in the convert_ANY_sublink_to_join() if VALUES
RTE is proper and the transformation is convertible. The conversion is only
possible for operations on scalar values, not rows. Also, we currently
support the transformation only when it ends up with a constant array.
Otherwise, the evaluation of non-hashed SAOP might be slower than the
corresponding Hash Join with VALUES.
Discussion: https://postgr.es/m/0184212d-1248-4f1f-a42d-f5cb1c1976d2%40tantorlabs.com
Author: Alena Rybakina <a.rybakina@postgrespro.ru>
Author: Andrei Lepikhov <lepihov@gmail.com>
Reviewed-by: Ivan Kush <ivan.kush@tantorlabs.com>
Reviewed-by: Alexander Korotkov <aekorotkov@gmail.com>
This commit extracts the code to generate ScalarArrayOpExpr on top of the list
of expressions from match_orclause_to_indexcol() into a separate function
make_SAOP_expr(). This function was extracted to be used in optimization for
conversion of 'x IN (VALUES ...)' to 'x = ANY ...'. make_SAOP_expr() is
placed in clauses.c file as only two additional headers were needed there
compared with other places.
Discussion: https://postgr.es/m/0184212d-1248-4f1f-a42d-f5cb1c1976d2%40tantorlabs.com
Author: Alena Rybakina <a.rybakina@postgrespro.ru>
Author: Andrei Lepikhov <lepihov@gmail.com>
Reviewed-by: Ivan Kush <ivan.kush@tantorlabs.com>
Reviewed-by: Alexander Korotkov <aekorotkov@gmail.com>
Previously, ALTER DEFAULT PRIVILEGES did not support large objects.
This meant that to grant privileges to users other than the owner,
permissions had to be manually assigned each time a large object
was created, which was inconvenient.
This commit extends ALTER DEFAULT PRIVILEGES to allow defining default
access privileges for large objects. With this change, specified privileges
will automatically apply to newly created large objects, making privilege
management more efficient.
As a side effect, this commit introduces the new keyword OBJECTS
since it's used in the syntax of ALTER DEFAULT PRIVILEGES.
Original patch by Haruka Takatsuka, with some fixes and tests by Yugo Nagata,
and rebased by Laurenz Albe.
Author: Takatsuka Haruka <harukat@sraoss.co.jp>
Co-authored-by: Yugo Nagata <nagata@sraoss.co.jp>
Co-authored-by: Laurenz Albe <laurenz.albe@cybertec.at>
Reviewed-by: Masao Fujii <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/20240424115242.236b499b2bed5b7a27f7a418@sraoss.co.jp
This gets rid of the bespoken ProcessWalRcvInterrupts() function,
which lets walreceiver terminate at any CHECK_FOR_INTERRUPTS() call.
And it's less code anyway.
We can now use the standard libpqsrv_connect_params() libpq wrapper
from libpq-be-fe-helpers.h, removing more code. We attempted to do
that earlier already in commit 728f86fec6, but that was reverted
because it didn't call ProcessWalRcvInterrupts() and therefore didn't
react to shutdown requests. Now that ProcessWalRcvInterrupts() is
gone, it works. As stated in that commit, this also leads to
libpqwalreceiver reserving file descriptors for libpq conncetions,
which is nice.
Author: Andres Freund <andres@anarazel.de> (the earlier commit)
Author: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Yura Sokolov <y.sokolov@postgrespro.ru>
Change the PathKey struct to use CompareType to record the sort
direction instead of hardcoding btree strategy numbers. The
CompareType is then converted to the index-type-specific strategy when
the plan is created.
This reduces the number of places btree strategy numbers are
hardcoded, and it's a self-contained subset of a larger effort to
allow non-btree indexes to behave like btrees.
Author: Mark Dilger <mark.dilger@enterprisedb.com>
Co-authored-by: Peter Eisentraut <peter@eisentraut.org>
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
This reverts commit f5930f9a98ea65d659d41600a138e608988ad122.
This broke the expansion of private hash tables, which reallocates the
directory. But that's impossible when it's allocated together with the
other fields, and dir_realloc() failed with BogusFree. Clearly, this
needs rethinking.
Discussion: https://postgr.es/m/CAApHDvriCiNkm=v521AP6PKPfyWkJ++jqZ9eqX4cXnhxLv8w-A@mail.gmail.com
Derived clauses are stored in ec_derives, a List of RestrictInfos.
These clauses are later looked up by matching the left and right
EquivalenceMembers along with the clause's parent EC.
This linear search becomes expensive in queries with many joins or
partitions, where ec_derives may contain thousands of entries. In
particular, create_join_clause() can spend significant time scanning
this list.
To improve performance, introduce a hash table (ec_derives_hash) that
is built when the list reaches 32 entries -- the same threshold used
for join_rel_hash. The original list is retained alongside the hash
table to support EC merging and serialization
(_outEquivalenceClass()).
Each clause is stored in the hash table using a canonicalized key: the
EquivalenceMember with the lower memory address is placed in the key
before the one with the higher memory address. This avoids storing or
searching for both permutations of the same clause. For clauses
involving a constant EM, the key places NULL in the first slot and the
non-constant EM in the second.
The hash table is initialized using list_length(ec_derives_list) as
the size hint. simplehash internally adjusts this to the next power of
two after dividing by the fillfactor, so this typically results in at
least 64 buckets near the threshold -- avoiding immediate resizing
while adapting to the actual number of entries.
The lookup logic for derived clauses is now centralized in
ec_search_derived_clause_for_ems(), which consults the hash table when
available and falls back to the list otherwise.
The new ec_clear_derived_clauses() always frees ec_derives_list, even
though some of the original code paths that cleared the old
ec_derives field did not. This ensures consistent cleanup and avoids
leaking memory when large lists are discarded.
An assertion originally placed in find_derived_clause_for_ec_member()
is moved into ec_search_derived_clause_for_ems() so that it is
enforced consistently, regardless of whether the hash table or list is
used for lookup.
This design incorporates suggestions by David Rowley, who proposed
both the key canonicalization and the initial sizing approach to
balance memory usage and CPU efficiency.
Author: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Reviewed-by: Amit Langote <amitlangote09@gmail.com>
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Tested-by: Dmitry Dolgov <9erthalion6@gmail.com>
Tested-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Tested-by: Amit Langote <amitlangote09@gmail.com>
Tested-by: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CAExHW5vZiQtWU6moszLP5iZ8gLX_ZAUbgEX0DxGLx9PGWCtqUg@mail.gmail.com
The issue is that the transactions prepared before two-phase decoding is
enabled can fail to replicate to the subscriber after being committed on a
promoted standby following a failover. This is because the two_phase_at
field of a slot, which tracks the LSN from which two-phase decoding
starts, is not synchronized to standby servers. Without two_phase_at, the
logical decoding might incorrectly identify prepared transaction as
already replicated to the subscriber after promotion of standby server,
causing them to be skipped.
To address the issue on HEAD, the two_phase_at field of the slot is
exposed by the pg_replication_slots view and allows the slot
synchronization to copy this value to the corresponding synced slot on the
standby server.
This bug is likely to occur if the user toggles the two_phase option to
true after initial slot creation. Given that altering the two_phase option
of a replication slot is not allowed in PostgreSQL 17, this bug is less
likely to occur. We can't change the view/function definition in
backbranch so we can't push the same fix but we are brainstorming an
appropriate solution for PG17.
Author: Zhijie Hou <houzj.fnst@fujitsu.com>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
Discussion: https://postgr.es/m/TYAPR01MB5724CC7C288535BBCEEE65DA94A72@TYAPR01MB5724.jpnprd01.prod.outlook.com
The optimization does not take the removal of TIDs by a concurrent vacuum into
account. The concurrent vacuum can remove dead TIDs and make pages ALL_VISIBLE
while those dead TIDs are referenced in the bitmap. This can lead to a
skip_fetch scan returning too many tuples.
It likely would be possible to implement this optimization safely, but we
don't have the necessary infrastructure in place. Nor is it clear that it's
worth building that infrastructure, given how limited the skip_fetch
optimization is.
In the backbranches we just disable the optimization by always passing
need_tuples=true to table_beginscan_bm(). We can't perform API/ABI changes in
the backbranches and we want to make the change as minimal as possible.
Author: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reported-By: Konstantin Knizhnik <knizhnik@garret.ru>
Discussion: https://postgr.es/m/CAEze2Wg3gXXZTr6_rwC+s4-o2ZVFB5F985uUSgJTsECx6AmGcQ@mail.gmail.com
Backpatch-through: 13
In the historical implementation of SQL functions (if they don't get
inlined), we built plans for all the contained queries at first call
within an outer query, and then re-used those plans for the duration
of the outer query, and then forgot everything. This was not ideal,
not least because the plans could not be customized to specific values
of the function's parameters. Our plancache infrastructure seems
mature enough to be used here. That will solve both the problem with
not being able to build custom plans and the problem with not being
able to share work across successive outer queries.
Aside from those performance concerns, this change fixes a
longstanding bugaboo with SQL functions: you could not write DDL that
would affect later statements in the same function. That's mostly
still true with new-style SQL functions, since the results of parse
analysis are baked into the stored query trees (and protected by
dependency records). But for old-style SQL functions, it will now
work much as it does with PL/pgSQL functions, because we delay parse
analysis and planning of each query until we're ready to run it.
Some edge cases that require replanning are now handled better too;
see for example the new rowsecurity test, where we now detect an RLS
context change that was previously missed.
One other edge-case change that might be worthy of a release note
is that we now insist that a SQL function's result be generated
by the physically-last query within it. Previously, if the last
original query was deleted by a DO INSTEAD NOTHING rule, we'd be
willing to take the result from the preceding query instead.
This behavior was undocumented except in source-code comments,
and it seems hard to believe that anyone's relying on it.
Along the way to this feature, we needed a few infrastructure changes:
* The plancache can now take either a raw parse tree or an
analyzed-but-not-rewritten Query as the starting point for a
CachedPlanSource. If given a Query, it is caller's responsibility
that nothing will happen to invalidate that form of the query.
We use this for new-style SQL functions, where what's in pg_proc is
serialized Query(s) and we trust the dependency mechanism to disallow
DDL that would break those.
* The plancache now offers a way to invoke a post-rewrite callback
to examine/modify the rewritten parse tree when it is rebuilding
the parse trees after a cache invalidation. We need this because
SQL functions sometimes adjust the parse tree to make its output
exactly match the declared result type; if the plan gets rebuilt,
that has to be re-done.
* There is a new backend module utils/cache/funccache.c that
abstracts the idea of caching data about a specific function
usage (a particular function and set of input data types).
The code in it is moved almost verbatim from PL/pgSQL, which
has done that for a long time. We use that logic now for
SQL-language functions too, and maybe other PLs will have use
for it in the future.
Author: Alexander Pyhalov <a.pyhalov@postgrespro.ru>
Co-authored-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Pavel Stehule <pavel.stehule@gmail.com>
Discussion: https://postgr.es/m/8216639.NyiUUSuA9g@aivenlaptop
For GiST, having a sortsupport function allows building the index
using the "sorted build" method, which is much faster.
For b-tree, the sortsupport routine doesn't give any new
functionality, but speeds up sorting a tiny bit. The difference is not
very significant, about 2% in cursory testing on my laptop, because
the range type comparison function has quite a lot of overhead from
detoasting. In any case, since we have the function for GiST anyway,
we might as well register it for the btree opfamily too.
Author: Bernd Helmle <mailings@oopsware.de>
Discussion: https://www.postgresql.org/message-id/64d324ce2a6d535d3f0f3baeeea7b25beff82ce4.camel@oopsware.de
pg_shmem_allocations tracks the memory allocated by ShmemInitStruct(),
but for shared hash tables that covered only the header and hash
directory. The remaining parts (segments and buckets) were allocated
later using ShmemAlloc(), which does not update the shmem accounting.
Thus, these allocations were not shown in pg_shmem_allocations.
This commit improves the situation by allocating all the hash table
parts at once, using a single ShmemInitStruct() call. This way the
ShmemIndex entries (and thus pg_shmem_allocations) better reflect the
proper size of the hash table.
This affects allocations for private (non-shared) hash tables too, as
the hash_create() code is shared. For non-shared tables this however
makes no practical difference.
This changes the alignment a bit. ShmemAlloc() aligns the chunks using
CACHELINEALIGN(), which means some parts (header, directory, segments)
were aligned this way. Allocating all parts as a single chunk removes
this (implicit) alignment. We've considered adding explicit alignment,
but we've decided not to - it seems to be merely a coincidence due to
using the ShmemAlloc() API, not due to necessity.
Author: Rahila Syed <rahilasyed90@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/CAH2L28vHzRankszhqz7deXURxKncxfirnuW68zD7+hVAqaS5GQ@mail.gmail.com
Currently, the cancel request key is a 32-bit token, which isn't very
much entropy. If you want to cancel another session's query, you can
brute-force it. In most environments, an unauthorized cancellation of
a query isn't very serious, but it nevertheless would be nice to have
more protection from it. Hence make the key longer, to make it harder
to guess.
The longer cancellation keys are generated when using the new protocol
version 3.2. For connections using version 3.0, short 4-bytes keys are
still used.
The new longer key length is not hardcoded in the protocol anymore,
the client is expected to deal with variable length keys, up to 256
bytes. This flexibility allows e.g. a connection pooler to add more
information to the cancel key, which might be useful for finding the
connection.
Reviewed-by: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Robert Haas <robertmhaas@gmail.com> (earlier versions)
Discussion: https://www.postgresql.org/message-id/508d0505-8b7a-4864-a681-e7e5edfe32aa@iki.fi
All supported version of the PostgreSQL server send the
NegotiateProtocolVersion message when an unsupported minor protocol
version is requested by a client. But many other applications that
implement the PostgreSQL protocol (connection poolers, or other
databases) do not, and the same is true for PostgreSQL server versions
older than 9.3. Connecting to such other applications thus fails if a
client requests a protocol version different than 3.0.
This patch adds a max_protocol_version connection option to libpq that
specifies the protocol version that libpq should request from the
server. Currently only 3.0 is supported, but that will change in a
future commit that bumps the protocol version. Even after that version
bump the default will likely stay 3.0 for the time being. Once more of
the ecosystem supports the NegotiateProtocolVersion message we might
want to change the default to the latest minor version.
This also adds the similar min_protocol_version connection option, to
allow the client to specify that connecting should fail if a lower
protocol version is attempted by the server. This can be used to
ensure that certain protocol features are used, which can be
particularly useful if those features impact security.
Author: Jelte Fennema-Nio <postgres@jeltef.nl>
Reviewed-by: Robert Haas <robertmhaas@gmail.com> (earlier versions)
Discussion: https://www.postgresql.org/message-id/CAGECzQTfc_O%2BHXqAo5_-xG4r3EFVsTefUeQzSvhEyyLDba-O9w@mail.gmail.com
Discussion: https://www.postgresql.org/message-id/CAGECzQRbAGqJnnJJxTdKewTsNOovUt4bsx3NFfofz3m2j-t7tA@mail.gmail.com
The changes made in commit d2b4b4c2259 contained incorrect comments:
They said that certain forward declarations were necessary to "avoid
including pathnodes.h here", but the file is itself pathnodes.h! So
change the comment to just say it's a forward declaration in one case,
and in the other case we don't need the declaration at all because it
already appeared earlier in the file.
timingsafe_bcmp() should be used instead of memcmp() or a naive
for-loop, when comparing passwords or secret tokens, to avoid leaking
information about the secret token by timing. This commit just
introduces the function but does not change any existing code to use
it yet.
Co-authored-by: Jelte Fennema-Nio <github-tech@jeltef.nl>
Discussion: https://www.postgresql.org/message-id/7b86da3b-9356-4e50-aa1b-56570825e234@iki.fi
This expands the NOT ENFORCED constraint flag, previously only
supported for CHECK constraints (commit ca87c415e2f), to foreign key
constraints.
Normally, when a foreign key constraint is created on a table, action
and check triggers are added to maintain data integrity. With this
patch, if a constraint is marked as NOT ENFORCED, integrity checks are
no longer required, making these triggers unnecessary. Consequently,
when creating a NOT ENFORCED foreign key constraint, triggers will not
be created, and the constraint will be marked as NOT VALID.
Similarly, if an existing foreign key constraint is changed to NOT
ENFORCED, the associated triggers will be dropped, and the constraint
will also be marked as NOT VALID. Conversely, if a NOT ENFORCED
foreign key constraint is changed to ENFORCED, the necessary triggers
will be created, and the will be changed to VALID by performing
necessary validation.
Since not-enforced foreign key constraints have no triggers, the
shortcut used for example in psql and pg_dump to skip looking for
foreign keys if the relation is known not to have triggers no longer
applies. (It already didn't work for partitioned tables.)
Author: Amul Sul <sulamul@gmail.com>
Reviewed-by: Joel Jacobson <joel@compiler.org>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Peter Eisentraut <peter@eisentraut.org>
Reviewed-by: jian he <jian.universality@gmail.com>
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Reviewed-by: Isaac Morland <isaac.morland@gmail.com>
Reviewed-by: Alexandra Wang <alexandra.wang.oss@gmail.com>
Tested-by: Triveni N <triveni.n@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/CAAJ_b962c5AcYW9KUt_R_ER5qs3fUGbe4az-SP-vuwPS-w-AGA@mail.gmail.com
Allow multiple backends to initialize WAL buffers concurrently. This way
`MemSet((char *) NewPage, 0, XLOG_BLCKSZ);` can run in parallel without
taking a single LWLock in exclusive mode.
The new algorithm works as follows:
* reserve a page for initialization using XLogCtl->InitializeReserved,
* ensure the page is written out,
* once the page is initialized, try to advance XLogCtl->InitializedUpTo and
signal to waiters using XLogCtl->InitializedUpToCondVar condition
variable,
* repeat previous steps until we reserve initialization up to the target
WAL position,
* wait until concurrent initialization finishes using a
XLogCtl->InitializedUpToCondVar.
Now, multiple backends can, in parallel, concurrently reserve pages,
initialize them, and advance XLogCtl->InitializedUpTo to point to the latest
initialized page.
Author: Yura Sokolov <y.sokolov@postgrespro.ru>
Co-authored-by: Alexander Korotkov <aekorotkov@gmail.com>
Reviewed-by: Pavel Borisov <pashkin.elfe@gmail.com>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Tested-by: Michael Paquier <michael@paquier.xyz>
Even after reaching the minimum recovery point, if there are long-lived
write transactions with 64 subtransactions on the primary, the recovery
snapshot may not yet be ready for hot standby, delaying read-only
connections on the standby. Previously, when read-only connections were
not accepted due to this condition, the following error message was logged:
FATAL: the database system is not yet accepting connections
DETAIL: Consistent recovery state has not been yet reached.
This DETAIL message was misleading because the following message was
already logged in this case:
LOG: consistent recovery state reached
This contradiction, i.e., indicating that the recovery state was consistent
while also stating it wasn’t, caused confusion.
This commit improves the error message to better reflect the actual state:
FATAL: the database system is not yet accepting connections
DETAIL: Recovery snapshot is not yet ready for hot standby.
HINT: To enable hot standby, close write transactions with more than 64 subtransactions on the primary server.
To implement this, the commit introduces a new postmaster signal,
PMSIGNAL_RECOVERY_CONSISTENT. When the startup process reaches
a consistent recovery state, it sends this signal to the postmaster,
allowing it to correctly recognize that state.
Since this is not a clear bug, the change is applied only to the master
branch and is not back-patched.
Author: Atsushi Torikoshi <torikoshia@oss.nttdata.com>
Co-authored-by: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Yugo Nagata <nagata@sraoss.co.jp>
Discussion: https://postgr.es/m/02db8cd8e1f527a8b999b94a4bee3165@oss.nttdata.com
As of 15.4, macOS has strchrnul(), but access to it is blocked behind
a check for MACOSX_DEPLOYMENT_TARGET >= 15.4. But our does-it-link
configure check finds it, so we try to use it, and fail with the
present default deployment target (namely 15.0). This accounts for
today's buildfarm failures on indri and sifaka.
This is the identical problem that we faced some years ago when Apple
introduced preadv and pwritev in the same way. We solved that in
commit f014b1b9b by using AC_CHECK_DECLS instead of AC_CHECK_FUNCS
to check the functions' availability. So do the same now for
strchrnul(). Interestingly, we already had a workaround for
"the link check doesn't agree with <string.h>" cases with glibc,
which we no longer need since only the header declaration is being
checked.
Testing this revealed that the meson version of this check has never
worked, because it failed to use "-Werror=unguarded-availability-new".
(Apparently nobody's tried to build with meson on macOS versions that
lack preadv/pwritev as standard.) Adjust that while at it. Also,
we had never put support for "-Werror=unguarded-availability-new"
into v13, but we need that now.
Co-authored-by: Tom Lane <tgl@sss.pgh.pa.us>
Co-authored-by: Peter Eisentraut <peter@eisentraut.org>
Discussion: https://postgr.es/m/385134.1743523038@sss.pgh.pa.us
Backpatch-through: 13
On Red Hat 9 systems (or similar), the packaged gcc targets x86-64-v2,
but clang does not. This has caused build failures in the wake of
commit e2809e3a1 when building --with-llvm.
The most expedient fix is to use the same function attributes for
the inlined function as we do for the global function.
Reported-by: Kyotaro Horiguchi <horikyota.ntt@gmail.com> (plus members skimmer and bumblebee)
Diagnosed-by: Tom Lane <tgl@sss.pgh.pa.us>
Tested-by: Todd Cook <cookt@blackduck.com>
Discussion: https://postgr.es/m/CANWCAZZSxs3a1YRKehkgk2OHKbrVn+xZ+AWW8Co2R_f70NqqmA@mail.gmail.com
pg_crc32c.h now has a simplified copy of the loop in pg_crc32c_sse42.c
suitable for inlining where possible.
This may slightly reduce contention for the WAL insertion lock,
but that hasn't been tested. The motivation for this change is avoid
regressing for a future commit that will use a function pointer for
non-constant input in all x86 builds.
While it's technically possible to make a similar change for Arm and
LoongArch, there are some questions about how inlining should work
since those platforms prefer stricter alignment. There are also no
immediate plans to add additional implementations for them.
Reviewed-by: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Raghuveer Devulapalli <raghuveer.devulapalli@intel.com>
Discussion: https://postgr.es/m/CANWCAZZEiTzhZcuwTiJ2=opiNpAUn1vuDRu1N02z61AthwRZLA@mail.gmail.com
Discussion: https://postgr.es/m/CANWCAZYRhLHArpyfV4uRK-Rw9N5oV5HMkkKtBehcuTjNOMwCZg@mail.gmail.com
Previously effective_io_concurrency and maintenance_io_concurrency could not
be set above 0 on machines without fadvise support. AIO enables IO concurrency
without such support, via io_method=worker.
Currently only subsystems using the read stream API will take advantage of
this. Other users of maintenance_io_concurrency (like recovery prefetching)
which leverage OS advice directly will not benefit from this change. In those
cases, maintenance_io_concurrency will have no effect on I/O behavior.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Discussion: https://postgr.es/m/CAAKRu_atGgZePo=_g6T3cNtfMf0QxpvoUh5OUqa_cnPdhLd=gw@mail.gmail.com
Submitting IO in larger batches can be more efficient than doing so
one-by-one, particularly for many small reads. It does, however, require
the ReadStreamBlockNumberCB callback to abide by the restrictions of AIO
batching (c.f. pgaio_enter_batchmode()). Basically, the callback may not:
a) block without first calling pgaio_submit_staged(), unless a
to-be-waited-on lock cannot be part of a deadlock, e.g. because it is
never held while waiting for IO.
b) directly or indirectly start another batch pgaio_enter_batchmode()
As this requires care and is nontrivial in some cases, batching is only
used with explicit opt-in.
This patch adds an explicit flag (READ_STREAM_USE_BATCHING) to read_stream and
uses it where appropriate.
There are two cases where batching would likely be beneficial, but where we
aren't using it yet:
1) bitmap heap scans, because the callback reads the VM
This should soon be solved, because we are planning to remove the use of
the VM, due to that not being sound.
2) The first phase of heap vacuum
This could be made to support batchmode, but would require some care.
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/uvrtrknj4kdytuboidbhwclo4gxhswwcpgadptsjvjqcluzmah%40brqs62irg4dt
This finally introduces the first actual use of AIO. StartReadBuffers() now
uses the AIO routines to issue IO.
As the implementation of StartReadBuffers() is also used by the functions for
reading individual blocks (StartReadBuffer() and through that
ReadBufferExtended()) this means all buffered read IO passes through the AIO
paths. However, as those are synchronous reads, actually performing the IO
asynchronously would be rarely beneficial. Instead such IOs are flagged to
always be executed synchronously. This way we don't have to duplicate a fair
bit of code.
When io_method=sync is used, the IO patterns generated after this change are
the same as before, i.e. actual reads are only issued in WaitReadBuffers() and
StartReadBuffers() may issue prefetch requests. This allows to bypass most of
the actual asynchronicity, which is important to make a change as big as this
less risky.
One thing worth calling out is that, if IO is actually executed
asynchronously, the precise meaning of what track_io_timing is measuring has
changed. Previously it tracked the time for each IO, but that does not make
sense when multiple IOs are executed concurrently. Now it only measures the
time actually spent waiting for IO. A subsequent commit will adjust the docs
for this.
While AIO is now actually used, the logic in read_stream.c will often prevent
using sufficiently many concurrent IOs. That will be addressed in the next
commit.
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Co-authored-by: Andres Freund <andres@anarazel.de>
Co-authored-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/uvrtrknj4kdytuboidbhwclo4gxhswwcpgadptsjvjqcluzmah%40brqs62irg4dt
Discussion: https://postgr.es/m/20210223100344.llw5an2aklengrmn@alap3.anarazel.de
Discussion: https://postgr.es/m/stj36ea6yyhoxtqkhpieia2z4krnam7qyetc57rfezgk4zgapf@gcnactj4z56m
This commit implements the infrastructure to perform asynchronous reads into
the buffer pool.
To do so, it:
- Adds readv AIO callbacks for shared and local buffers
It may be worth calling out that shared buffer completions may be run in a
different backend than where the IO started.
- Adds an AIO wait reference to BufferDesc, to allow backends to wait for
in-progress asynchronous IOs
- Adapts StartBufferIO(), WaitIO(), TerminateBufferIO(), and their localbuf.c
equivalents, to be able to deal with AIO
- Moves the code to handle BM_PIN_COUNT_WAITER into a helper function, as it
now also needs to be called on IO completion
As of this commit, nothing issues AIO on shared/local buffers. A future commit
will update StartReadBuffers() to do so.
Buffer reads executed through this infrastructure will report invalid page /
checksum errors / warnings differently than before:
In the error case the error message will cover all the blocks that were
included in the read, rather than just the reporting the first invalid
block. If more than one block is invalid, the error will include information
about the range of the read, the first invalid block and the number of invalid
pages, with a HINT towards the server log for per-block details.
For the warning case (i.e. zero_damaged_buffers) we would previously emit one
warning message for each buffer in a multi-block read. Now there is only a
single warning message for the entire read, again referring to the server log
for more details in case of multiple checksum failures within a single larger
read.
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/uvrtrknj4kdytuboidbhwclo4gxhswwcpgadptsjvjqcluzmah%40brqs62irg4dt
Discussion: https://postgr.es/m/20210223100344.llw5an2aklengrmn@alap3.anarazel.de
Discussion: https://postgr.es/m/stj36ea6yyhoxtqkhpieia2z4krnam7qyetc57rfezgk4zgapf@gcnactj4z56m
If an IO succeeds, but issues a warning, e.g. due to a page verification
failure with zero_damaged_pages, we want to issue that warning in the context
of the issuer of the IO, not the process that executes the completion (always
the case for worker).
It's already possible for a completion callback to report a custom error
message, we just didn't have a result status that allowed a user of AIO to
know that a warning should be emitted even though the IO request succeeded.
All that's needed for that is a dedicated PGAIO_RS_ value.
Previously there were not enough bits in PgAioResult.id for the new
value. Increase. While at that, add defines for the amount of bits and static
asserts to check that the widths are appropriate.
Reviewed-by: Noah Misch <noah@leadboat.com>
Discussion: https://postgr.es/m/20250329212929.a6.nmisch@google.com
For AIO the completion of a read into shared buffers (i.e. verifying the page
including the checksum, updating the BufferDesc to reflect the IO) can happen
in a different backend than the backend that started the IO. As
ignore_checksum_failure can differ between backends, we need to allow the
caller of PageIsVerified() control whether to ignore checksum failures.
The commit leaves a gap in the PIV_* values, as an upcoming commit, which
depends on this commit, will add PIV_LOG_LOG, which better fits just after
PIV_LOG_WARNING.
Reviewed-by: Noah Misch <noah@leadboat.com>
Discussion: https://postgr.es/m/20250329212929.a6.nmisch@google.com
