Until now BufferDesc.state was not allowed to be modified while the buffer
header spinlock was held. This meant that operations like unpinning buffers
needed to use a CAS loop, waiting for the buffer header spinlock to be
released before updating.
The benefit of that restriction is that it allowed us to unlock the buffer
header spinlock with just a write barrier and an unlocked write (instead of a
full atomic operation). That was important to avoid regressions in
48354581a4. However, since then the hottest buffer header spinlock uses have
been replaced with atomic operations (in particular, the most common use of
PinBuffer_Locked(), in GetVictimBuffer() (formerly in BufferAlloc()), has been
removed in 5e89985928).
This change will allow, in a subsequent commit, to release buffer pins with a
single atomic-sub operation. This previously was not possible while such
operations were not allowed while the buffer header spinlock was held, as an
atomic-sub would not have allowed a race-free check for the buffer header lock
being held.
Using atomic-sub to unpin buffers is a nice scalability win, however it is not
the primary motivation for this change (although it would be sufficient). The
primary motivation is that we would like to merge the buffer content lock into
BufferDesc.state, which will result in more frequent changes of the state
variable, which in some situations can cause a performance regression, due to
an increased CAS failure rate when unpinning buffers. The regression entirely
vanishes when using atomic-sub.
Naively implementing this would require putting CAS loops in every place
modifying the buffer state while holding the buffer header lock. To avoid
that, introduce UnlockBufHdrExt(), which can set/add flags as well as the
refcount, together with releasing the lock.
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://postgr.es/m/fvfmkr5kk4nyex56ejgxj3uzi63isfxovp2biecb4bspbjrze7@az2pljabhnff
Previously StrategyGetBuffer() acquired the buffer header spinlock for every
buffer, whether it was reusable or not. If reusable, it'd be returned, with
the lock held, to GetVictimBuffer(), which then would pin the buffer with
PinBuffer_Locked(). That's somewhat violating the spirit of the guidelines for
holding spinlocks (i.e. that they are only held for a few lines of consecutive
code) and necessitates using PinBuffer_Locked(), which scales worse than
PinBuffer() due to holding the spinlock. This alone makes it worth changing
the code.
However, the main reason to change this is that a future commit will make
PinBuffer_Locked() slower (due to making UnlockBufHdr() slower), to gain
scalability for the much more common case of pinning a pre-existing buffer. By
pinning the buffer with a single atomic operation, iff the buffer is reusable,
we avoid any potential regression for miss-heavy workloads. There strictly are
fewer atomic operations for each potential buffer after this change.
The price for this improvement is that freelist.c needs two CAS loops and
needs to be able to set up the resource accounting for pinned buffers. The
latter is achieved by exposing a new function for that purpose from bufmgr.c,
that seems better than exposing the entire private refcount infrastructure.
The improvement seems worth the complexity.
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://postgr.es/m/fvfmkr5kk4nyex56ejgxj3uzi63isfxovp2biecb4bspbjrze7@az2pljabhnff
This set of changes removes the list of available buffers and instead simply
uses the clock-sweep algorithm to find and return an available buffer. This
also removes the have_free_buffer() function and simply caps the
pg_autoprewarm process to at most NBuffers.
While on the surface this appears to be removing an optimization it is in fact
eliminating code that induces overhead in the form of synchronization that is
problematic for multi-core systems.
The main reason for removing the freelist, however, is not the moderate
improvement in scalability, but that having the freelist would require
dedicated complexity in several upcoming patches. As we have not been able to
find a case benefiting from the freelist...
Author: Greg Burd <greg@burd.me>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/70C6A5B5-2A20-4D0B-BC73-EB09DD62D61C@getmailspring.com
Before, BAS_BULKREAD was always of size 256kB. With the default
io_combine_limit of 16, that only allowed 1-2 IOs to be in flight -
insufficient even on very low latency storage.
We don't just want to increase the size to a much larger hardcoded value, as
very large rings (10s of MBs of of buffers), appear to have negative
performance effects when reading in data that the OS has cached (but not when
actually needing to do IO).
To address this, increase the size of BAS_BULKREAD to allow for
io_combine_limit * effective_io_concurrency buffers getting read in. To
prevent the ring being much larger than useful, limit the increased size with
GetPinLimit().
The formula outlined above keeps the ring size to sizes for which we have not
observed performance regressions, unless very large effective_io_concurrency
values are used together with large shared_buffers setting.
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/lqwghabtu2ak4wknzycufqjm5ijnxhb4k73vzphlt2a3wsemcd@gtftg44kdim6
Discussion: https://postgr.es/m/uvrtrknj4kdytuboidbhwclo4gxhswwcpgadptsjvjqcluzmah@brqs62irg4dt
To make them follow the usual naming convention where
FoobarShmemSize() calculates the amount of shared memory needed by
Foobar subsystem, and FoobarShmemInit() performs the initialization.
I didn't rename CreateLWLocks() and InitShmmeIndex(), because they are
a little special. They need to be called before any of the other
ShmemInit() functions, because they set up the shared memory
bookkeeping itself. I also didn't rename InitProcGlobal(), because
unlike other Shmeminit functions, it's not called by individual
backends.
Reviewed-by: Andreas Karlsson
Discussion: https://www.postgresql.org/message-id/c09694ff-2453-47e5-b26c-32a16cd75ce6@iki.fi
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
Add new options to the VACUUM and ANALYZE commands called
BUFFER_USAGE_LIMIT to allow users more control over how large to make the
buffer access strategy that is used to limit the usage of buffers in
shared buffers. Larger rings can allow VACUUM to run more quickly but
have the drawback of VACUUM possibly evicting more buffers from shared
buffers that might be useful for other queries running on the database.
Here we also add a new GUC named vacuum_buffer_usage_limit which controls
how large to make the access strategy when it's not specified in the
VACUUM/ANALYZE command. This defaults to 256KB, which is the same size as
the access strategy was prior to this change. This setting also
controls how large to make the buffer access strategy for autovacuum.
Per idea by Andres Freund.
Author: Melanie Plageman
Reviewed-by: David Rowley
Reviewed-by: Andres Freund
Reviewed-by: Justin Pryzby
Reviewed-by: Bharath Rupireddy
Discussion: https://postgr.es/m/20230111182720.ejifsclfwymw2reb@awork3.anarazel.de
The new name better reflects what the field is - the size of the buffers[]
array. ring_size sounded more like it is in units of bytes.
An upcoming commit allows a BufferAccessStrategy of custom sizes, so it
seems relevant to improve this beforehand.
Author: Melanie Plageman
Reviewed-by: David Rowley
Discussion: https://postgr.es/m/CAAKRu_YefVbhg4rAxU2frYFdTap61UftH-kUNQZBvAs%2BYK81Zg%40mail.gmail.com
Commit 28e626bde0 introduced the infrastructure for tracking more detailed IO
statistics. This commit adds the actual collection of the new IO statistics
for relations and temporary relations. See aforementioned commit for goals and
high-level design.
The changes in this commit are fairly straight-forward. The bulk of the change
is to passing sufficient information to the callsites of pgstat_count_io_op().
A somewhat unsightly detail is that it currently is hard to find a better
place to count fsyncs than in md.c, whereas the other pgstat_count_io_op()
calls are in bufmgr.c/localbuf.c. As the number of fsyncs is tied to md.c
implementation details, it's not obvious there is a better answer.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20200124195226.lth52iydq2n2uilq@alap3.anarazel.de
Periodically while the server is running, and at shutdown, write out a
list of blocks in shared buffers. When the server reaches consistency
-- unfortunatey, we can't do it before that point without breaking
things -- reload those blocks into any still-unused shared buffers.
Mithun Cy and Robert Haas, reviewed and tested by Beena Emerson,
Amit Kapila, Jim Nasby, and Rafia Sabih.
Discussion: http://postgr.es/m/CAD__OugubOs1Vy7kgF6xTjmEqTR4CrGAv8w+ZbaY_+MZeitukw@mail.gmail.com
Don't move parenthesized lines to the left, even if that means they
flow past the right margin.
By default, BSD indent lines up statement continuation lines that are
within parentheses so that they start just to the right of the preceding
left parenthesis. However, traditionally, if that resulted in the
continuation line extending to the right of the desired right margin,
then indent would push it left just far enough to not overrun the margin,
if it could do so without making the continuation line start to the left of
the current statement indent. That makes for a weird mix of indentations
unless one has been completely rigid about never violating the 80-column
limit.
This behavior has been pretty universally panned by Postgres developers.
Hence, disable it with indent's new -lpl switch, so that parenthesized
lines are always lined up with the preceding left paren.
This patch is much less interesting than the first round of indent
changes, but also bulkier, so I thought it best to separate the effects.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
The new indent version includes numerous fixes thanks to Piotr Stefaniak.
The main changes visible in this commit are:
* Nicer formatting of function-pointer declarations.
* No longer unexpectedly removes spaces in expressions using casts,
sizeof, or offsetof.
* No longer wants to add a space in "struct structname *varname", as
well as some similar cases for const- or volatile-qualified pointers.
* Declarations using PG_USED_FOR_ASSERTS_ONLY are formatted more nicely.
* Fixes bug where comments following declarations were sometimes placed
with no space separating them from the code.
* Fixes some odd decisions for comments following case labels.
* Fixes some cases where comments following code were indented to less
than the expected column 33.
On the less good side, it now tends to put more whitespace around typedef
names that are not listed in typedefs.list. This might encourage us to
put more effort into typedef name collection; it's not really a bug in
indent itself.
There are more changes coming after this round, having to do with comment
indentation and alignment of lines appearing within parentheses. I wanted
to limit the size of the diffs to something that could be reviewed without
one's eyes completely glazing over, so it seemed better to split up the
changes as much as practical.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
Pinning/Unpinning a buffer is a very frequent operation; especially in
read-mostly cache resident workloads. Benchmarking shows that in various
scenarios the spinlock protecting a buffer header's state becomes a
significant bottleneck. The problem can be reproduced with pgbench -S on
larger machines, but can be considerably worse for queries which touch
the same buffers over and over at a high frequency (e.g. nested loops
over a small inner table).
To allow atomic operations to be used, cram BufferDesc's flags,
usage_count, buf_hdr_lock, refcount into a single 32bit atomic variable;
that allows to manipulate them together using 32bit compare-and-swap
operations. This requires reducing MAX_BACKENDS to 2^18-1 (which could
be lifted by using a 64bit field, but it's not a realistic configuration
atm).
As not all operations can easily implemented in a lockfree manner,
implement the previous buf_hdr_lock via a flag bit in the atomic
variable. That way we can continue to lock the header in places where
it's needed, but can get away without acquiring it in the more frequent
hot-paths. There's some additional operations which can be done without
the lock, but aren't in this patch; but the most important places are
covered.
As bufmgr.c now essentially re-implements spinlocks, abstract the delay
logic from s_lock.c into something more generic. It now has already two
users, and more are coming up; there's a follupw patch for lwlock.c at
least.
This patch is based on a proof-of-concept written by me, which Alexander
Korotkov made into a fully working patch; the committed version is again
revised by me. Benchmarking and testing has, amongst others, been
provided by Dilip Kumar, Alexander Korotkov, Robert Haas.
On a large x86 system improvements for readonly pgbench, with a high
client count, of a factor of 8 have been observed.
Author: Alexander Korotkov and Andres Freund
Discussion: 2400449.GjM57CE0Yg@dinodell
Prior to commit 0709b7ee72, access to
variables within a spinlock-protected critical section had to be done
through a volatile pointer, but that should no longer be necessary.
Review by Andres Freund
Use "a" and "an" correctly, mostly in comments. Two error messages were
also fixed (they were just elogs, so no translation work required). Two
function comments in pg_proc.h were also fixed. Etsuro Fujita reported one
of these, but I found a lot more with grep.
Also fix a few other typos spotted while grepping for the a/an typos.
For example, "consists out of ..." -> "consists of ...". Plus a "though"/
"through" mixup reported by Euler Taveira.
Many of these typos were in old code, which would be nice to backpatch to
make future backpatching easier. But much of the code was new, and I didn't
feel like crafting separate patches for each branch. So no backpatching.
Benchmarks has shown that aligning the buffer descriptor array to
cache lines is important for scalability; especially on bigger,
multi-socket, machines.
Currently the array sometimes already happens to be aligned by
happenstance, depending how large previous shared memory allocations
were. That can lead to wildly varying performance results after minor
configuration changes.
In addition to aligning the start of descriptor array, also force the
size of individual descriptors to be of a common cache line size (64
bytes). That happens to already be the case on 64bit platforms, but
this way we can change the struct BufferDesc more easily.
As the alignment primarily matters in highly concurrent workloads
which probably all are 64bit these days, and the space wastage of
element alignment would be a bit more noticeable on 32bit systems, we
don't force the stride to be cacheline sized on 32bit platforms for
now. If somebody does actual performance testing, we can reevaluate
that decision by changing the definition of BUFFERDESC_PADDED_SIZE.
Discussion: 20140202151319.GD32123@awork2.anarazel.de
Per discussion with Bruce Momjan, Tom Lane, Robert Haas, and Peter
Geoghegan.
Since their introduction latches have required barriers in SetLatch
and ResetLatch - but when they were introduced there wasn't any
barrier abstraction. Instead latches were documented to rely on the
callsites to provide barrier semantics.
Now that the barrier support looks halfway complete, add the necessary
barriers to both latch implementations.
Also remove a now superflous lock acquisition from syncrep.c and a
superflous (and insufficient) barrier from freelist.c. There might be
other cases that can now be simplified, but those are the only ones
I've seen on a quick scan.
We might want to backpatch this at some later point, but right now the
barrier infrastructure in the backbranches isn't totally on par with
master.
Discussion: 20150112154026.GB2092@awork2.anarazel.de
StrategyGetBuffer() has proven to be a bottleneck in a number of
buffer acquisition heavy workloads. To some degree this has already
been alleviated by 5d7962c6, but it still can be quite a heavy
bottleneck. The problem is that in unfortunate usage patterns a
single StrategyGetBuffer() call will have to look at a large number of
buffers - in turn making it likely that the process will be put to
sleep while still holding the spinlock.
Replace most of the usage of the buffer_strategy_lock spinlock for the
clock sweep by a atomic nextVictimBuffer variable. That variable,
modulo NBuffers, is the current hand of the clock sweep. The buffer
clock-sweep then only needs to acquire the spinlock after a
wraparound. And even then only in the process that did the wrapping
around. That alleviates nearly all the contention on the relevant
spinlock, although significant contention on the cacheline can still
exist.
Reviewed-By: Robert Haas and Amit Kapila
Discussion: 20141010160020.GG6670@alap3.anarazel.de,
20141027133218.GA2639@awork2.anarazel.de
Previously, we used an lwlock that was held from the time we began
seeking a candidate buffer until the time when we found and pinned
one, which is disastrous for concurrency. Instead, use a spinlock
which is held just long enough to pop the freelist or advance the
clock sweep hand, and then released. If we need to advance the clock
sweep further, we reacquire the spinlock once per buffer.
This represents a significant increase in atomic operations around
buffer eviction, but it still wins on many workloads. On others, it
may result in no gain, or even cause a regression, unless the number
of buffer mapping locks is also increased. However, that seems like
material for a separate commit. We may also need to consider other
methods of mitigating contention on this spinlock, such as splitting
it into multiple locks or jumping the clock sweep hand more than one
buffer at a time, but those, too, seem like separate improvements.
Patch by me, inspired by a much larger patch from Amit Kapila.
Reviewed by Andres Freund.
The Solaris Studio compiler warns about these instances, unlike more
mainstream compilers such as gcc. But manual inspection showed that
the code is clearly not reachable, and we hope no worthy compiler will
complain about removing this code.
Commit 6d90eaaa89 added a hibernation mode
to the bgwriter to reduce the server's idle-power consumption. However,
its interaction with the detailed behavior of BgBufferSync's feedback
control loop wasn't very well thought out. That control loop depends
primarily on the rate of buffer allocation, not the rate of buffer
dirtying, so the hibernation mode has to be designed to operate only when
no new buffer allocations are happening. Also, the check for whether the
system is effectively idle was not quite right and would fail to detect
a constant low level of activity, thus allowing the bgwriter to go into
hibernation mode in a way that would let the cycle time vary quite a bit,
possibly further confusing the feedback loop. To fix, move the wakeup
support from MarkBufferDirty and SetBufferCommitInfoNeedsSave into
StrategyGetBuffer, and prevent the bgwriter from entering hibernation mode
unless no buffer allocations have happened recently.
In addition, fix the delaying logic to remove the problem of possibly not
responding to signals promptly, which was basically caused by trying to use
the process latch's is_set flag for multiple purposes. I can't prove it
but I'm suspicious that that hack was responsible for the intermittent
"postmaster does not shut down" failures we've been seeing in the buildfarm
lately. In any case it did nothing to improve the readability or
robustness of the code.
In passing, express the hibernation sleep time as a multiplier on
BgWriterDelay, not a constant. I'm not sure whether there's any value in
exposing the longer sleep time as an independently configurable setting,
but we can at least make it act like this for little extra code.
