29f6a959c added a bump allocator type for efficient compact allocations.
Here we make use of this for non-bounded tuplesorts to store tuples.
This is very space efficient when storing narrow tuples due to bump.c
not having chunk headers. This means we can fit more tuples in work_mem
before spilling to disk, or perform an in-memory sort touching fewer
cacheline.
Author: David Rowley
Reviewed-by: Nathan Bossart
Reviewed-by: Matthias van de Meent
Reviewed-by: Tomas Vondra
Reviewed-by: John Naylor
Discussion: https://postgr.es/m/CAApHDvqGSpCU95TmM=Bp=6xjL_nLys4zdZOpfNyWBk97Xrdj2w@mail.gmail.com
This tracks the position of WAL that's been fully copied into WAL
buffers by all processes emitting WAL. (For some reason we call that
"WAL insertion"). This is updated using atomic monotonic advance during
WaitXLogInsertionsToFinish, which is not when the insertions actually
occur, but it's the only place where we know where have all the
insertions have completed.
This value is useful in WALReadFromBuffers, which can verify that
callers don't try to read past what has been inserted. (However, more
infrastructure is needed in order to actually use WAL after the flush
point, since it could be lost.)
The value is also useful in WaitXLogInsertionsToFinish() itself, since
we can now exit quickly when all WAL has been already inserted, without
even having to take any locks.
This introduces a bump MemoryContext type. The bump context is best
suited for short-lived memory contexts which require only allocations
of memory and never a pfree or repalloc, which are unsupported.
Memory palloc'd into a bump context has no chunk header. This makes
bump a useful context type when lots of small allocations need to be
done without any need to pfree those allocations. Allocation sizes are
rounded up to the next MAXALIGN boundary, so with this and no chunk
header, allocations are very compact indeed.
Allocations are also very fast as bump does not check any freelists to
try and make use of previously free'd chunks. It just checks if there
is enough room on the current block, and if so it bumps the freeptr
beyond this chunk and returns the value that the freeptr was previously
pointing to. Simple and fast. A new block is malloc'd when there's not
enough space in the current block.
Code using the bump allocator must take care never to call any functions
which could try to call realloc() (or variants), pfree(),
GetMemoryChunkContext() or GetMemoryChunkSpace() on a bump allocated
chunk. Due to lack of chunk headers, these operations are unsupported.
To increase the chances of catching such issues, when compiled with
MEMORY_CONTEXT_CHECKING, bump allocated chunks are given a header and
any attempt to perform an unsupported operation will result in an ERROR.
Without MEMORY_CONTEXT_CHECKING, code attempting an unsupported
operation could result in a segfault.
A follow-on commit will implement the first user of bump.
Author: David Rowley
Reviewed-by: Nathan Bossart
Reviewed-by: Matthias van de Meent
Reviewed-by: Tomas Vondra
Reviewed-by: John Naylor
Discussion: https://postgr.es/m/CAApHDvqGSpCU95TmM=Bp=6xjL_nLys4zdZOpfNyWBk97Xrdj2w@mail.gmail.com
Reserve 4 bits for MemoryContextMethodID rather than 3. 3 bits did
technically allow a maximum of 8 memory context types, however, we've
opted to reserve some bit patterns which left us with only 4 slots, all
of which were used.
Here we add another bit which frees up 8 slots for future memory context
types.
In passing, adjust the enum names in MemoryContextMethodID to make it
more clear which ones can be used and which ones are reserved.
Author: Matthias van de Meent, David Rowley
Discussion: https://postgr.es/m/CAApHDvqGSpCU95TmM=Bp=6xjL_nLys4zdZOpfNyWBk97Xrdj2w@mail.gmail.com
Until now, when calling pq_putmessage to write new data to a libpq
socket, all writes are copied into a buffer and that buffer gets flushed
when full to avoid having to perform small writes to the socket.
There are cases where we must write large amounts of data to the socket,
sometimes larger than the size of the buffer. In this case, it's
wasteful to memcpy this data into the buffer and flush it out, instead,
we can send it directly from the memory location that the data is already
stored in.
Here we adjust internal_putbytes() so that after having just flushed the
buffer to the socket, if the remaining bytes to send is as big or bigger
than the buffer size, we just send directly rather than needlessly
copying into the PqSendBuffer buffer first.
Examples of operations that write large amounts of data in one message
are; outputting large tuples with SELECT or COPY TO STDOUT and
pg_basebackup.
Author: Melih Mutlu
Reviewed-by: Heikki Linnakangas
Reviewed-by: Jelte Fennema-Nio
Reviewed-by: David Rowley
Reviewed-by: Ranier Vilela
Reviewed-by: Andres Freund
Discussion: https://postgr.es/m/CAGPVpCR15nosj0f6xe-c2h477zFR88q12e6WjEoEZc8ZYkTh3Q@mail.gmail.com
Until now, heapgetpage()'s loop over all tuples performed some conditional
checks for each tuple, even though condition did not change across the loop.
This commit fixes that by moving the loop into an inline function. By calling
it with different constant arguments, the compiler can generate an optimized
loop for the different conditions, at the price of two per-page checks.
For cases of all-visible tables and an isolation level other than
serializable, speedups of up to 25% have been measured.
Reviewed-by: John Naylor <johncnaylorls@gmail.com>
Reviewed-by: Zhang Mingli <zmlpostgres@gmail.com>
Tested-by: Quan Zongliang <quanzongliang@yeah.net>
Discussion: https://postgr.es/m/20230716015656.xjvemfbp5fysjiea@awork3.anarazel.de
Discussion: https://postgr.es/m/2ef7ff1b-3d18-2283-61b1-bbd25fc6c7ce@yeah.net
Commit 792752af4e added infrastructure for using AVX-512 intrinsic
functions, and this commit uses that infrastructure to optimize
visibilitymap_count(). Specificially, a new pg_popcount_masked()
function is introduced that applies a bitmask to every byte in the
buffer prior to calculating the population count, which is used to
filter out the all-visible or all-frozen bits as needed. Platforms
without AVX-512 support should also see a nice speedup due to the
reduced number of calls to a function pointer.
Co-authored-by: Ants Aasma
Discussion: https://postgr.es/m/BL1PR11MB5304097DF7EA81D04C33F3D1DCA6A%40BL1PR11MB5304.namprd11.prod.outlook.com
Presently, pg_popcount() processes data in 32-bit or 64-bit chunks
when possible. Newer hardware that supports AVX-512 instructions
can use 512-bit chunks, which provides a nice speedup, especially
for larger buffers. This commit introduces the infrastructure
required to detect compiler and CPU support for the required
AVX-512 intrinsic functions, and it adds a new pg_popcount()
implementation that uses these functions. If CPU support for this
optimized implementation is detected at runtime, a function pointer
is updated so that it is used by subsequent calls to pg_popcount().
Most of the existing in-tree calls to pg_popcount() should benefit
from these instructions, and calls with smaller buffers should at
least not regress compared to v16. The new infrastructure
introduced by this commit can also be used to optimize
visibilitymap_count(), but that is left for a follow-up commit.
Co-authored-by: Paul Amonson, Ants Aasma
Reviewed-by: Matthias van de Meent, Tom Lane, Noah Misch, Akash Shankaran, Alvaro Herrera, Andres Freund, David Rowley
Discussion: https://postgr.es/m/BL1PR11MB5304097DF7EA81D04C33F3D1DCA6A%40BL1PR11MB5304.namprd11.prod.outlook.com
When we determine that a wanted block can't be combined with the current
pending read, it's time to start that read to get it out of the way. An
"if" in that code path should have been a "while", because it might take
more than one go in case of partial reads. This was only broken for
smaller ranges, as the more common case of io_combine_limit-sized ranges
is handled earlier in the code and knows how to loop, hiding the bug for
a while.
Discovered while testing large parallel sequential scans of partially
cached tables. The ramp-up-and-down block allocator for parallel scans
could hit the problem case and skip some blocks near the end that should
have been streamed.
Defect in commit b5a9b18c.
Discussion: https://postgr.es/m/CA%2BhUKG%2Bh8Whpv0YsJqjMVkjYX%2B80fTVc6oi-V%2BzxJvykLpLHYQ%40mail.gmail.com
The buildfarm members using debug_parallel_query = regress are mostly
unhappy with this test. I guess what is happening is that rows
generated by a parallel worker are buffered, and might or might not
get to the leader before the expected error occurs. We did not see
any variability in the old version of this test because each FETCH
would succeed or fail atomically, leading to a predictable number of
rows emitted before failure. I don't find this to be a bug, just
unspecified behavior, so let's disable parallel query for this one
test case to make the results stable.
Formerly this was done with a cursor, which is problematic since
not all result-set-returning query types can be put into a cursor.
The new implementation is better integrated into other psql
features, too.
Daniel Vérité, reviewed by Laurenz Albe and myself (and whacked
around a bit by me, so any remaining bugs are my fault)
Discussion: https://postgr.es/m/CAKZiRmxsVTkO928CM+-ADvsMyePmU3L9DQCa9NwqjvLPcEe5QA@mail.gmail.com
This patch generalizes libpq's existing single-row mode to allow
individual partial-result PGresults to contain up to N rows, rather
than always one row. This reduces malloc overhead compared to plain
single-row mode, and it is very useful for psql's FETCH_COUNT feature,
since otherwise we'd have to add code (and cycles) to either merge
single-row PGresults into a bigger one or teach psql's
results-printing logic to accept arrays of PGresults.
To avoid API breakage, PQsetSingleRowMode() remains the same, and we
add a new function PQsetChunkedRowsMode() to invoke the more general
case. Also, PGresults obtained the old way continue to carry the
PGRES_SINGLE_TUPLE status code, while if PQsetChunkedRowsMode() is
used then their status code is PGRES_TUPLES_CHUNK. The underlying
logic is the same either way, though.
Daniel Vérité, reviewed by Laurenz Albe and myself (and whacked
around a bit by me, so any remaining bugs are my fault)
Discussion: https://postgr.es/m/CAKZiRmxsVTkO928CM+-ADvsMyePmU3L9DQCa9NwqjvLPcEe5QA@mail.gmail.com
BitmapAdjustPrefetchIterator() only used the blockno member of the
passed in TBMIterateResult to ensure that the prefetch iterator and
regular iterator stay in sync. Pass it the BlockNumber only, so that we
can move away from using the TBMIterateResult outside of table AM
specific code.
Author: Melanie Plageman
Reviewed-by: Tomas Vondra, Andres Freund, Heikki Linnakangas
Discussion: https://postgr.es/m/CAAKRu_ZwCwWFeL_H3ia26bP2e7HiKLWt0ZmGXPVwPO6uXq0vaA%40mail.gmail.com
As of 7c70996ebf, BitmapPrefetch() used the recheck flag for
the current block to determine whether or not it should skip prefetching
the proposed prefetch block. As explained in the comment, this assumed
the index AM will report the same recheck value for the future page as
it did for the current page - but there's no guarantee.
This only affects prefetching - if the recheck flag changes, we may
prefetch blocks unecessarily and not prefetch blocks that will be
needed. But we don't need to rely on that assumption - we know the
recheck flag for the block we're considering prefetching, so we can
use that.
The impact is very limited in practice - the opclass would need to
assign different recheck flags to different blocks, but none of the
built-in opclasses seems to do that.
Author: Melanie Plageman
Reviewed-by: Tomas Vondra, Andres Freund, Tom Lane
Discussion: https://postgr.es/m/1939305.1712415547%40sss.pgh.pa.us
Commit 7c70996ebf introduced an optimization to allow bitmap
scans to operate like index-only scans by not fetching a block from the
heap if none of the underlying data is needed and the block is marked
all visible in the visibility map.
With the introduction of table AMs, a FIXME was added to this code
indicating that the skip_fetch logic should be pushed into the table
AM-specific code, as not all table AMs may use a visibility map in the
same way.
This commit resolves this FIXME for the current block. The layering
violation is still present in BitmapHeapScans's prefetching code, which
uses the visibility map to decide whether or not to prefetch a block.
However, this can be addressed independently.
Author: Melanie Plageman
Reviewed-by: Andres Freund, Heikki Linnakangas, Tomas Vondra, Mark Dilger
Discussion: https://postgr.es/m/CAAKRu_ZwCwWFeL_H3ia26bP2e7HiKLWt0ZmGXPVwPO6uXq0vaA%40mail.gmail.com
This new DDL command splits a single partition into several parititions.
Just like ALTER TABLE ... MERGE PARTITIONS ... command, new patitions are
created using createPartitionTable() function with parent partition as the
template.
This commit comprises quite naive implementation which works in single process
and holds the ACCESS EXCLUSIVE LOCK on the parent table during all the
operations including the tuple routing. This is why this new DDL command
can't be recommended for large partitioned tables under a high load. However,
this implementation come in handy in certain cases even as is.
Also, it could be used as a foundation for future implementations with lesser
locking and possibly parallel.
Discussion: https://postgr.es/m/c73a1746-0cd0-6bdd-6b23-3ae0b7c0c582%40postgrespro.ru
Author: Dmitry Koval
Reviewed-by: Matthias van de Meent, Laurenz Albe, Zhihong Yu, Justin Pryzby
Reviewed-by: Alvaro Herrera, Robert Haas, Stephane Tachoires
This new DDL command merges several partitions into the one partition of the
target table. The target partition is created using new
createPartitionTable() function with parent partition as the template.
This commit comprises quite naive implementation which works in single process
and holds the ACCESS EXCLUSIVE LOCK on the parent table during all the
operations including the tuple routing. This is why this new DDL command
can't be recommended for large partitioned tables under a high load. However,
this implementation come in handy in certain cases even as is.
Also, it could be used as a foundation for future implementations with lesser
locking and possibly parallel.
Discussion: https://postgr.es/m/c73a1746-0cd0-6bdd-6b23-3ae0b7c0c582%40postgrespro.ru
Author: Dmitry Koval
Reviewed-by: Matthias van de Meent, Laurenz Albe, Zhihong Yu, Justin Pryzby
Reviewed-by: Alvaro Herrera, Robert Haas, Stephane Tachoires
Set BitmapHeapScanState->can_skip_fetch in BitmapHeapNext() instead of
in ExecInitBitmapHeapScan(). This is a preliminary step to pushing the
skip fetch optimization into heap AM code.
Author: Melanie Plageman
Reviewed-by: Tomas Vondra, Andres Freund, Heikki Linnakangas
Discussion: https://postgr.es/m/CAAKRu_ZwCwWFeL_H3ia26bP2e7HiKLWt0ZmGXPVwPO6uXq0vaA%40mail.gmail.com
Even though waiting for replay LSN happens without explicit transaction,
AbortTransaction() is responsible for the cleanup of the shared memory if
the error is thrown in a stored procedure. So, we need to do WaitLSNCleanup()
there to clean up after some unexpected error happened while waiting for
replay LSN.
Discussion: https://postgr.es/m/202404051815.eri4u5q6oj26%40alvherre.pgsql
Author: Alvaro Herrera
Not just WaitLSNState.waitersHeap, but also WaitLSNState.procInfos and
updating of WaitLSNState.minWaitedLSN is protected by WaitLSNLock. There
is one now documented exclusion on fast-path checking of
WaitLSNProcInfo.inHeap flag.
Discussion: https://postgr.es/m/202404030658.hhj3vfxeyhft%40alvherre.pgsql
Change the order so that the table scan is initialized only after
initializing the index scan and building the bitmap.
This is mostly a cosmetic change for now, but later commits will need
to pass parameters to table_beginscan_bm() that are unavailable in
ExecInitBitmapHeapScan().
Author: Melanie Plageman
Reviewed-by: Tomas Vondra, Andres Freund, Heikki Linnakangas
Discussion: https://postgr.es/m/CAAKRu_ZwCwWFeL_H3ia26bP2e7HiKLWt0ZmGXPVwPO6uXq0vaA%40mail.gmail.com
Commit 9e8da0f7 taught nbtree to handle ScalarArrayOpExpr quals
natively. This works by pushing down the full context (the array keys)
to the nbtree index AM, enabling it to execute multiple primitive index
scans that the planner treats as one continuous index scan/index path.
This earlier enhancement enabled nbtree ScalarArrayOp index-only scans.
It also allowed scans with ScalarArrayOp quals to return ordered results
(with some notable restrictions, described further down).
Take this general approach a lot further: teach nbtree SAOP index scans
to decide how to execute ScalarArrayOp scans (when and where to start
the next primitive index scan) based on physical index characteristics.
This can be far more efficient. All SAOP scans will now reliably avoid
duplicative leaf page accesses (just like any other nbtree index scan).
SAOP scans whose array keys are naturally clustered together now require
far fewer index descents, since we'll reliably avoid starting a new
primitive scan just to get to a later offset from the same leaf page.
The scan's arrays now advance using binary searches for the array
element that best matches the next tuple's attribute value. Required
scan key arrays (i.e. arrays from scan keys that can terminate the scan)
ratchet forward in lockstep with the index scan. Non-required arrays
(i.e. arrays from scan keys that can only exclude non-matching tuples)
"advance" without the process ever rolling over to a higher-order array.
Naturally, only required SAOP scan keys trigger skipping over leaf pages
(non-required arrays cannot safely end or start primitive index scans).
Consequently, even index scans of a composite index with a high-order
inequality scan key (which we'll mark required) and a low-order SAOP
scan key (which we won't mark required) now avoid repeating leaf page
accesses -- that benefit isn't limited to simpler equality-only cases.
In general, all nbtree index scans now output tuples as if they were one
continuous index scan -- even scans that mix a high-order inequality
with lower-order SAOP equalities reliably output tuples in index order.
This allows us to remove a couple of special cases that were applied
when building index paths with SAOP clauses during planning.
Bugfix commit 807a40c5 taught the planner to avoid generating unsafe
path keys: path keys on a multicolumn index path, with a SAOP clause on
any attribute beyond the first/most significant attribute. These cases
are now all safe, so we go back to generating path keys without regard
for the presence of SAOP clauses (just like with any other clause type).
Affected queries can now exploit scan output order in all the usual ways
(e.g., certain "ORDER BY ... LIMIT n" queries can now terminate early).
Also undo changes from follow-up bugfix commit a4523c5a, which taught
the planner to produce alternative index paths, with path keys, but
without low-order SAOP index quals (filter quals were used instead).
We'll no longer generate these alternative paths, since they can no
longer offer any meaningful advantages over standard index qual paths.
Affected queries thereby avoid all of the disadvantages that come from
using filter quals within index scan nodes. They can avoid extra heap
page accesses from using filter quals to exclude non-matching tuples
(index quals will never have that problem). They can also skip over
irrelevant sections of the index in more cases (though only when nbtree
determines that starting another primitive scan actually makes sense).
There is a theoretical risk that removing restrictions on SAOP index
paths from the planner will break compatibility with amcanorder-based
index AMs maintained as extensions. Such an index AM could have the
same limitations around ordered SAOP scans as nbtree had up until now.
Adding a pro forma incompatibility item about the issue to the Postgres
17 release notes seems like a good idea.
Author: Peter Geoghegan <pg@bowt.ie>
Author: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reviewed-By: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-By: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reviewed-By: Tomas Vondra <tomas.vondra@enterprisedb.com>
Discussion: https://postgr.es/m/CAH2-Wz=ksvN_sjcnD1+Bt-WtifRA5ok48aDYnq3pkKhxgMQpcw@mail.gmail.com
When this bit of commentary was written, it was alluding to the
fact that we looked for newlines and EOD markers in the raw
(not yet encoding-converted) input data. We don't do that anymore,
preferring to batch the conversion of larger chunks of input and
split it into lines later. Hence there's no longer any need for
assumptions about the relevant characters being encoding-invariant,
and we should remove this comment saying we assume that.
Discussion: https://postgr.es/m/1461688.1712347668@sss.pgh.pa.us
After encountering the NUL terminator, the word-at-a-time loop exits
and we must hash the remaining bytes. Previously we calculated
the terminator's position and re-loaded the remaining bytes from
the input string. This was slower than the unaligned case for very
short strings. We already have all the data we need in a register,
so let's just mask off the bytes we need and hash them immediately.
In addition to endianness issues, the previous attempt upset valgrind
in the way it computed the mask. Whether by accident or by wisdom,
the author's proposed method passes locally with valgrind 3.22.
Ants Aasma, with cosmetic adjustments by me
Discussion: https://postgr.es/m/CANwKhkP7pCiW_5fAswLhs71-JKGEz1c1%2BPC0a_w1fwY4iGMqUA%40mail.gmail.com
This function previously used a mix of word-wise loads and bytewise
loads. The bytewise loads happened to be little-endian regardless of
platform. This in itself is not a problem. However, a future commit
will require the same result whether A) the input is loaded as a
word with the relevent bytes masked-off, or B) the input is loaded
one byte at a time.
While at it, improve debuggability of the internal hash state.
Discussion: https://postgr.es/m/CANWCAZZpuV1mES1mtSpAq8tWJewbrv4gEz6R_k4gzNG8GZ5gag%40mail.gmail.com
The BAS_VACUUM ring size has been 256kB since commit d526575f introduced
the mechanism 17 years ago. Commit 1cbbee03 recently made it
configurable but retained the traditional default. The correct default
size has been debated for years, but 256kB is certainly very small.
VACUUM soon needs to write back data it dirtied only 32 blocks ago,
which usually requires flushing the WAL. New experiments in prefetching
pages for VACUUM exacerbated the problem by crashing into dirty data
even sooner. Let's make the default 2MB. That's 1.6% of the default
toy buffer pool size, and 0.2% of 1GB, which would be a considered a
small shared_buffers setting for a real system these days. Users are
still free to set the GUC to a different value.
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20240403221257.md4gfki3z75cdyf6%40awork3.anarazel.de
Discussion: https://postgr.es/m/CA%2BhUKGLY4Q4ZY4f1rvnFtv6%2BPkjNf8MejdPkcju3Qii9DYqqcQ%40mail.gmail.com
While pinning extra buffers to look ahead, users of strategies are in
danger of using too many buffers. For some strategies, that means
"escaping" from the ring, and in others it means forcing dirty data to
disk very frequently with associated WAL flushing. Since external code
has no insight into any of that, allow individual strategy types to
expose a clamp that should be applied when deciding how many buffers to
pin at once.
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://postgr.es/m/CAAKRu_aJXnqsyZt6HwFLnxYEBgE17oypkxbKbT1t1geE_wvH2Q%40mail.gmail.com
Remove duplicate hash_string_pointer() function definitions by creating
a new inline function hash_string() for this purpose.
This has the added advantage of avoiding strlen() calls when doing hash
lookup. It's not clear how many of these are perfomance-sensitive
enough to benefit from that, but the simplification is worth it on
its own.
Reviewed by Jeff Davis
Discussion: https://postgr.es/m/CANWCAZbg_XeSeY0a_PqWmWqeRATvzTzUNYRLeT%2Bbzs%2BYQdC92g%40mail.gmail.com
fasthash_accum_cstring_aligned() uses a technique, found in various
strlen() implementations, to detect a string's NUL terminator by
reading a word at at time. That triggers failures when testing with
"-fsanitize=address", at least with frontend code. To enable using
this function anywhere, add a function attribute macro to disable
such testing.
Reviewed by Jeff Davis
Discussion: https://postgr.es/m/CANWCAZbwvp7oUEkbw-xP4L0_S_WNKq-J-ucP4RCNDPJnrakUPw%40mail.gmail.com
The "fast path" for well cached scans that don't do any I/O was
accidentally coded in a way that could only be triggered by pg_prewarm's
usage pattern, which starts out with a higher distance because of the
flags it passes in. We want it to work for streaming sequential scans
too, once that patch is committed. Adjust.
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://postgr.es/m/CA%2BhUKGKXZALJ%3D6aArUsXRJzBm%3Dqvc4AWp7%3DiJNXJQqpbRLnD_w%40mail.gmail.com
The previous formula was incorrect in the case where the function's
nblocks argument was a multiple of BLOCKS_PER_CHUNK, which happens
whenever a relation segment file is exactly 512MB or exactly 1GB in
length. In such cases, the formula would calculate a stop_offset of
0 rather than 65536, resulting in modified blocks in the second half
of a 1GB file, or all the modified blocks in a 512MB file, being
omitted from the incremental backup.
Reported off-list by Tomas Vondra and Jakub Wartak.
Discussion: http://postgr.es/m/CA+TgmoYwy_KHp1-5GYNmVa=zdeJWhNH1T0SBmEuvqQNJEHj1Lw@mail.gmail.com
Fix a mistake in ac81101551 - write_reconstructed_file() called
copy_file_range() without properly checking HAVE_COPY_FILE_RANGE.
Reported by several macOS machines. Also reported by cfbot, but I missed
that issue before commit.
This commit allows using copy_file_range() for efficient combining of
data from multiple files, instead of simply reading/writing the blocks.
Depending on the filesystem and other factors (size of the increment,
distribution of modified blocks etc.) this may be faster than the
block-by-block copy, but more importantly it enables various features
provided by CoW filesystems.
If a checksum needs to be calculated for the file, the same strategy as
when copying whole files is used - copy_file_range is used to copy the
blocks, but the file is also read for the checksum calculation.
While the checksum calculation is rarely needed when cloning whole
files, when reconstructing the files from multiple backups it needs to
happen almost always (the only exception is when the user specified
--no-manifest).
Author: Tomas Vondra
Reviewed-by: Thomas Munro, Jakub Wartak, Robert Haas
Discussion: https://postgr.es/m/3024283a-7491-4240-80d0-421575f6bb23%40enterprisedb.com
Adds --clone/--copy-file-range options to pg_combinebackup, to allow
copying files using file cloning or copy_file_range(). These methods may
be faster than the standard block-by-block copy, but the main advantage
is that they enable various features provided by CoW filesystems.
This commit only uses these copy methods for files that did not change
and can be copied as a whole from a single backup.
These new copy methods may not be available on all platforms, in which
case the command throws an error (immediately, even if no files would be
copied as a whole). This early failure seems better than failing later
when trying to copy the first file, after performing a lot of work on
earlier files.
If the requested copy method is available, but a checksum needs to be
recalculated (e.g. because of a different checksum type), the file is
still copied using the requested method, but it is also read for the
checksum calculation. Depending on the filesystem this may be more
expensive than just performing the simple copy, but it does enable the
CoW benefits.
Initial patch by Jakub Wartak, various reworks and improvements by me.
Author: Tomas Vondra, Jakub Wartak
Reviewed-by: Thomas Munro, Jakub Wartak, Robert Haas
Discussion: https://postgr.es/m/3024283a-7491-4240-80d0-421575f6bb23%40enterprisedb.com
Buildfarm member caiman is showing this, which surprises me because
it's very late-model gcc (14.0.1) and ought to be smart enough to
know that elog(ERROR) doesn't return. But we're likely to see the
same from stupider compilers too, so add a dummy initialization in
our usual style.
The documentation index is getting very long, which makes it hard
to find things. Since these chapters are all very similar in structure
and content, merging them is a natural way of reducing the size of
the toplevel index.
Rather than actually combining all of the SGML into a single file,
keep one file per <sect1>, and add a glue file that includes all
of them.
Discussion: http://postgr.es/m/CA+Tgmob7_uoYuS2=rVwpVXaRwP-UXz+++saYTC-BCZ42QzSNKQ@mail.gmail.com
Align blocks stored in incremental files to BLCKSZ, so that the
incremental backups work well with CoW filesystems.
The header of the incremental file is padded with \0 to a multiple of
BLCKSZ, so that the block data (also BLCKSZ) is aligned to BLCKSZ. The
padding is added only to files containing block data, so files with just
the header remain small. This adds a bit of extra space, but as the
number of blocks increases the overhead gets negligible very quickly.
And as the padding is \0 bytes, it does compress extremely well.
The alignment is important for CoW filesystems that usually require the
blocks to be aligned to filesystem page size for features like block
sharing, deduplication etc. to work well. With the variable sized header
the blocks in the increments were not aligned at all, negating the
benefits of the CoW filesystems.
This matters even for non-CoW filesystems, for example when placed on a
RAID array. If the block is not aligned, it may easily span multiple
devices, causing read and write amplification.
It might be better to align the blocks to the filesystem page, not
BLCKSZ, but we have no good way to determine that. Even if we determine
the page size at the time of taking the backup, the backup may move. For
now the BLCKSZ seems sufficient - the filesystem page is usually 4K, so
the default BLCKSZ (8K by default) is aligned to that.
Author: Tomas Vondra
Reviewed-by: Robert Haas, Jakub Wartak
Discussion: https://postgr.es/m/3024283a-7491-4240-80d0-421575f6bb23%40enterprisedb.com
This removes the need to hold both the info_lck spinlock and
WALWriteLock to update them. We use stock atomic write instead, with
WALWriteLock held. Readers can use atomic read, without any locking.
This allows for some code to be reordered: some places were a bit
contorted to avoid repeated spinlock acquisition, but that's no longer a
concern, so we can turn them to more natural coding. Some further
changes are possible (maybe to performance wins), but in this commit I
did rather minimal ones only, to avoid increasing the blast radius.
Reviewed-by: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Reviewed-by: Jeff Davis <pgsql@j-davis.com>
Reviewed-by: Andres Freund <andres@anarazel.de> (earlier versions)
Discussion: https://postgr.es/m/20200831182156.GA3983@alvherre.pgsql
This commit does two things:
1) Maintains inactive_since for sync slots whenever the slot is released
just like any other regular slot.
2) Ensures the value is set to the current timestamp during the promotion
of standby to help correctly interpret the time after promotion. We don't
want the slots to appear inactive for a long time after promotion if they
haven't been synchronized recently. This would also avoid the invalidation
of such slots immediately after promotion if tomorrow we have a feature
that invalidates slots based on their inactivity time. Whoever acquires
the slot i.e. makes the slot active will reset it to NULL.
Author: Bharath Rupireddy
Reviewed-by: Bertrand Drouvot, Amit Kapila, Shveta Malik, Masahiko Sawada
Discussion: https://postgr.es/m/CAA4eK1KrPGwfZV9LYGidjxHeW+rxJ=E2ThjXvwRGLO=iLNuo=Q@mail.gmail.com
Discussion: https://postgr.es/m/CALj2ACW4aUe-_uFQOjdWCEN-xXoLGhmvRFnL8SNw_TZ5nJe+aw@mail.gmail.com
Discussion: https://postgr.es/m/CA+Tgmob_Ta-t2ty8QrKHBGnNLrf4ZYcwhGHGFsuUoFrAEDw4sA@mail.gmail.com
The current design behind the automatic generation of the C code and
documentation related to wait events introduced in fa88928470 does not
offer a way to attach new wait events without breaking ABI
compatibility, as all the events are forcibly reordered for each section
in the input file wait_event_names.txt. Adding new wait events to
stable branches is something that has happened in the past, 0b6517a3b7
being a recent example of that with VERSION_FILE_SYNC, so we need a way
to generate any C code for wait events while maintaining compatibility
on stable branches already released.
This commit solves this issue by adding a new region called
"ABI_compatibility" (keyword could be updated to something else if
someone had a better idea) to each section of wait_event_names.txt, so
as one can add new wait events to stable branches in
wait_event_names.txt while keeping the code ABI-compatible.
"ABI_compatibility" has no impact on the documentation generated: all
the wait events of one section are still alphabetically ordered. LWLock
and Lock sections generate their C code elsewhere, so they do not need
an "ABI_compatibility" region.
For example, let's imagine a wait_event_names.txt like that:
Section: ClassName - Foo
FOO_1 "Waiting in Foo 1"
FOO_2 "Waiting in Foo 2"
ABI_compatibility:
NEW_FOO_1 "Waiting in New Foo 1"
NEW_BAR_1 "Waiting in New Bar 1"
This results in the following enum, where the events in the ABI region
are listed last with the same ordering as in wait_event_names.txt:
typedef enum
{
WAIT_EVENT_FOO_1,
WAIT_EVENT_FOO_2,
WAIT_EVENT_NEW_FOO_1,
WAIT_EVENT_NEW_BAR_1
} WaitEventFoo;
New wait events added in stable branches should be added at the end of
each ABI_compatibility region, and ABI_compatibility should remain empty
on HEAD and unreleased stable branches.
This design has been suggested by Noah Misch and me.
Reported-by: Noah Misch
Author: Bertrand Drouvot
Reviewed-by: Michael Paquier
Discussion: https://postgr.es/m/20240317183114.16@rfd.leadboat.com
David Rowley