On 7/25/25 15:02, Greg Burd wrote:
> Patch set is now:
>
> 1) remove freelist
>
> 2) remove buffer_strategy_lock
>
> 3) abstract clock-sweep to type and API
>
>
>
> -greg
Somehow including the test.c file as an attachment on my last email
confused the CI and it didn't test the v10 patch set (which did pass in
GitHub CI on my fork [1]). Here's v11 unchanged from v10 except rebased
onto 258bf0a2ea8 cf PG 19-2 5928 [2].
best,
-greg
[1] https://github.com/gburd/postgres/pull/10/checks
[2] https://commitfest.postgresql.org/patch/5928/
From 706dcef951f4ec33b492ace481d31aacb02fd650 Mon Sep 17 00:00:00 2001
From: Greg Burd <g...@burd.me>
Date: Thu, 10 Jul 2025 14:45:32 -0400
Subject: [PATCH v11 1/3] Eliminate the freelist from the buffer manager and
depend on clock-sweep
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. 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 problemmatic for multi-core systems.
This also removes the have_free_buffer() function and simply caps the
pg_autoprewarm process to at most NBuffers.
---
contrib/pg_prewarm/autoprewarm.c | 31 ++++---
src/backend/storage/buffer/README | 42 +++------
src/backend/storage/buffer/buf_init.c | 9 --
src/backend/storage/buffer/bufmgr.c | 29 +------
src/backend/storage/buffer/freelist.c | 120 +++-----------------------
src/include/storage/buf_internals.h | 12 +--
6 files changed, 43 insertions(+), 200 deletions(-)
diff --git a/contrib/pg_prewarm/autoprewarm.c b/contrib/pg_prewarm/autoprewarm.c
index c01b9c7e6a4..2722b0bb443 100644
--- a/contrib/pg_prewarm/autoprewarm.c
+++ b/contrib/pg_prewarm/autoprewarm.c
@@ -370,6 +370,16 @@ apw_load_buffers(void)
apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx = 0;
apw_state->prewarmed_blocks = 0;
+
+ /* Don't prewarm more than we can fit. */
+ if (num_elements > NBuffers)
+ {
+ num_elements = NBuffers;
+ ereport(LOG,
+ (errmsg("autoprewarm: capping prewarmed blocks to %d (shared_buffers size)",
+ NBuffers)));
+ }
+
/* Get the info position of the first block of the next database. */
while (apw_state->prewarm_start_idx < num_elements)
{
@@ -410,10 +420,6 @@ apw_load_buffers(void)
apw_state->database = current_db;
Assert(apw_state->prewarm_start_idx < apw_state->prewarm_stop_idx);
- /* If we've run out of free buffers, don't launch another worker. */
- if (!have_free_buffer())
- break;
-
/*
* Likewise, don't launch if we've already been told to shut down.
* (The launch would fail anyway, but we might as well skip it.)
@@ -462,12 +468,6 @@ apw_read_stream_next_block(ReadStream *stream,
{
BlockInfoRecord blk = p->block_info[p->pos];
- if (!have_free_buffer())
- {
- p->pos = apw_state->prewarm_stop_idx;
- return InvalidBlockNumber;
- }
-
if (blk.tablespace != p->tablespace)
return InvalidBlockNumber;
@@ -523,10 +523,10 @@ autoprewarm_database_main(Datum main_arg)
blk = block_info[i];
/*
- * Loop until we run out of blocks to prewarm or until we run out of free
+ * Loop until we run out of blocks to prewarm or until we run out of
* buffers.
*/
- while (i < apw_state->prewarm_stop_idx && have_free_buffer())
+ while (i < apw_state->prewarm_stop_idx)
{
Oid tablespace = blk.tablespace;
RelFileNumber filenumber = blk.filenumber;
@@ -568,14 +568,13 @@ autoprewarm_database_main(Datum main_arg)
/*
* We have a relation; now let's loop until we find a valid fork of
- * the relation or we run out of free buffers. Once we've read from
- * all valid forks or run out of options, we'll close the relation and
+ * the relation or we run out of buffers. Once we've read from all
+ * valid forks or run out of options, we'll close the relation and
* move on.
*/
while (i < apw_state->prewarm_stop_idx &&
blk.tablespace == tablespace &&
- blk.filenumber == filenumber &&
- have_free_buffer())
+ blk.filenumber == filenumber)
{
ForkNumber forknum = blk.forknum;
BlockNumber nblocks;
diff --git a/src/backend/storage/buffer/README b/src/backend/storage/buffer/README
index a182fcd660c..cd52effd911 100644
--- a/src/backend/storage/buffer/README
+++ b/src/backend/storage/buffer/README
@@ -128,11 +128,11 @@ independently. If it is necessary to lock more than one partition at a time,
they must be locked in partition-number order to avoid risk of deadlock.
* A separate system-wide spinlock, buffer_strategy_lock, provides mutual
-exclusion for operations that access the buffer free list or select
-buffers for replacement. A spinlock is used here rather than a lightweight
-lock for efficiency; no other locks of any sort should be acquired while
-buffer_strategy_lock is held. This is essential to allow buffer replacement
-to happen in multiple backends with reasonable concurrency.
+exclusion for operations that select buffers for replacement. A spinlock is
+used here rather than a lightweight lock for efficiency; no other locks of any
+sort should be acquired while buffer_strategy_lock is held. This is essential
+to allow buffer replacement to happen in multiple backends with reasonable
+concurrency.
* Each buffer header contains a spinlock that must be taken when examining
or changing fields of that buffer header. This allows operations such as
@@ -158,18 +158,9 @@ unset by sleeping on the buffer's condition variable.
Normal Buffer Replacement Strategy
----------------------------------
-There is a "free list" of buffers that are prime candidates for replacement.
-In particular, buffers that are completely free (contain no valid page) are
-always in this list. We could also throw buffers into this list if we
-consider their pages unlikely to be needed soon; however, the current
-algorithm never does that. The list is singly-linked using fields in the
-buffer headers; we maintain head and tail pointers in global variables.
-(Note: although the list links are in the buffer headers, they are
-considered to be protected by the buffer_strategy_lock, not the buffer-header
-spinlocks.) To choose a victim buffer to recycle when there are no free
-buffers available, we use a simple clock-sweep algorithm, which avoids the
-need to take system-wide locks during common operations. It works like
-this:
+To choose a victim buffer to recycle when there are no free buffers available,
+we use a simple clock-sweep algorithm, which avoids the need to take
+system-wide locks during common operations. It works like this:
Each buffer header contains a usage counter, which is incremented (up to a
small limit value) whenever the buffer is pinned. (This requires only the
@@ -184,20 +175,14 @@ The algorithm for a process that needs to obtain a victim buffer is:
1. Obtain buffer_strategy_lock.
-2. If buffer free list is nonempty, remove its head buffer. Release
-buffer_strategy_lock. If the buffer is pinned or has a nonzero usage count,
-it cannot be used; ignore it go back to step 1. Otherwise, pin the buffer,
-and return it.
+2. Select the buffer pointed to by nextVictimBuffer, and circularly advance
+nextVictimBuffer for next time. Release buffer_strategy_lock.
-3. Otherwise, the buffer free list is empty. Select the buffer pointed to by
-nextVictimBuffer, and circularly advance nextVictimBuffer for next time.
-Release buffer_strategy_lock.
-
-4. If the selected buffer is pinned or has a nonzero usage count, it cannot
+3. If the selected buffer is pinned or has a nonzero usage count, it cannot
be used. Decrement its usage count (if nonzero), reacquire
buffer_strategy_lock, and return to step 3 to examine the next buffer.
-5. Pin the selected buffer, and return.
+4. Pin the selected buffer, and return.
(Note that if the selected buffer is dirty, we will have to write it out
before we can recycle it; if someone else pins the buffer meanwhile we will
@@ -234,7 +219,7 @@ the ring strategy effectively degrades to the normal strategy.
VACUUM uses a ring like sequential scans, however, the size of this ring is
controlled by the vacuum_buffer_usage_limit GUC. Dirty pages are not removed
-from the ring. Instead, WAL is flushed if needed to allow reuse of the
+from the ring. Instead, the WAL is flushed if needed to allow reuse of the
buffers. Before introducing the buffer ring strategy in 8.3, VACUUM's buffers
were sent to the freelist, which was effectively a buffer ring of 1 buffer,
resulting in excessive WAL flushing.
@@ -277,3 +262,4 @@ As of 8.4, background writer starts during recovery mode when there is
some form of potentially extended recovery to perform. It performs an
identical service to normal processing, except that checkpoints it
writes are technically restartpoints.
+
diff --git a/src/backend/storage/buffer/buf_init.c b/src/backend/storage/buffer/buf_init.c
index ed1dc488a42..6fd3a6bbac5 100644
--- a/src/backend/storage/buffer/buf_init.c
+++ b/src/backend/storage/buffer/buf_init.c
@@ -128,20 +128,11 @@ BufferManagerShmemInit(void)
pgaio_wref_clear(&buf->io_wref);
- /*
- * Initially link all the buffers together as unused. Subsequent
- * management of this list is done by freelist.c.
- */
- buf->freeNext = i + 1;
-
LWLockInitialize(BufferDescriptorGetContentLock(buf),
LWTRANCHE_BUFFER_CONTENT);
ConditionVariableInit(BufferDescriptorGetIOCV(buf));
}
-
- /* Correct last entry of linked list */
- GetBufferDescriptor(NBuffers - 1)->freeNext = FREENEXT_END_OF_LIST;
}
/* Init other shared buffer-management stuff */
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index 6afdd28dba6..af5ef025229 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -2099,12 +2099,6 @@ BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
*/
UnpinBuffer(victim_buf_hdr);
- /*
- * The victim buffer we acquired previously is clean and unused, let
- * it be found again quickly
- */
- StrategyFreeBuffer(victim_buf_hdr);
-
/* remaining code should match code at top of routine */
existing_buf_hdr = GetBufferDescriptor(existing_buf_id);
@@ -2163,8 +2157,7 @@ BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
}
/*
- * InvalidateBuffer -- mark a shared buffer invalid and return it to the
- * freelist.
+ * InvalidateBuffer -- mark a shared buffer invalid.
*
* The buffer header spinlock must be held at entry. We drop it before
* returning. (This is sane because the caller must have locked the
@@ -2262,11 +2255,6 @@ retry:
* Done with mapping lock.
*/
LWLockRelease(oldPartitionLock);
-
- /*
- * Insert the buffer at the head of the list of free buffers.
- */
- StrategyFreeBuffer(buf);
}
/*
@@ -2684,11 +2672,6 @@ ExtendBufferedRelShared(BufferManagerRelation bmr,
{
BufferDesc *buf_hdr = GetBufferDescriptor(buffers[i] - 1);
- /*
- * The victim buffer we acquired previously is clean and unused,
- * let it be found again quickly
- */
- StrategyFreeBuffer(buf_hdr);
UnpinBuffer(buf_hdr);
}
@@ -2763,12 +2746,6 @@ ExtendBufferedRelShared(BufferManagerRelation bmr,
valid = PinBuffer(existing_hdr, strategy);
LWLockRelease(partition_lock);
-
- /*
- * The victim buffer we acquired previously is clean and unused,
- * let it be found again quickly
- */
- StrategyFreeBuffer(victim_buf_hdr);
UnpinBuffer(victim_buf_hdr);
buffers[i] = BufferDescriptorGetBuffer(existing_hdr);
@@ -3666,8 +3643,8 @@ BgBufferSync(WritebackContext *wb_context)
uint32 new_recent_alloc;
/*
- * Find out where the freelist clock sweep currently is, and how many
- * buffer allocations have happened since our last call.
+ * Find out where the clock sweep currently is, and how many buffer
+ * allocations have happened since our last call.
*/
strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
diff --git a/src/backend/storage/buffer/freelist.c b/src/backend/storage/buffer/freelist.c
index 01909be0272..162c140fb9d 100644
--- a/src/backend/storage/buffer/freelist.c
+++ b/src/backend/storage/buffer/freelist.c
@@ -39,14 +39,6 @@ typedef struct
*/
pg_atomic_uint32 nextVictimBuffer;
- int firstFreeBuffer; /* Head of list of unused buffers */
- int lastFreeBuffer; /* Tail of list of unused buffers */
-
- /*
- * NOTE: lastFreeBuffer is undefined when firstFreeBuffer is -1 (that is,
- * when the list is empty)
- */
-
/*
* Statistics. These counters should be wide enough that they can't
* overflow during a single bgwriter cycle.
@@ -164,17 +156,16 @@ ClockSweepTick(void)
}
/*
- * have_free_buffer -- a lockless check to see if there is a free buffer in
- * buffer pool.
+ * have_free_buffer -- check if we've filled the buffer pool at startup
*
- * If the result is true that will become stale once free buffers are moved out
- * by other operations, so the caller who strictly want to use a free buffer
- * should not call this.
+ * Used exclusively by autoprewarm.
*/
bool
have_free_buffer(void)
{
- if (StrategyControl->firstFreeBuffer >= 0)
+ uint64 hand = pg_atomic_read_u64(&StrategyControl->nextVictimBuffer);
+
+ if (hand < NBuffers)
return true;
else
return false;
@@ -243,75 +234,14 @@ StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_r
}
/*
- * We count buffer allocation requests so that the bgwriter can estimate
- * the rate of buffer consumption. Note that buffers recycled by a
- * strategy object are intentionally not counted here.
+ * We keep an approximate count of buffer allocation requests so that the
+ * bgwriter can estimate the rate of buffer consumption. Note that
+ * buffers recycled by a strategy object are intentionally not counted
+ * here.
*/
pg_atomic_fetch_add_u32(&StrategyControl->numBufferAllocs, 1);
- /*
- * First check, without acquiring the lock, whether there's buffers in the
- * freelist. Since we otherwise don't require the spinlock in every
- * StrategyGetBuffer() invocation, it'd be sad to acquire it here -
- * uselessly in most cases. That obviously leaves a race where a buffer is
- * put on the freelist but we don't see the store yet - but that's pretty
- * harmless, it'll just get used during the next buffer acquisition.
- *
- * If there's buffers on the freelist, acquire the spinlock to pop one
- * buffer of the freelist. Then check whether that buffer is usable and
- * repeat if not.
- *
- * Note that the freeNext fields are considered to be protected by the
- * buffer_strategy_lock not the individual buffer spinlocks, so it's OK to
- * manipulate them without holding the spinlock.
- */
- if (StrategyControl->firstFreeBuffer >= 0)
- {
- while (true)
- {
- /* Acquire the spinlock to remove element from the freelist */
- SpinLockAcquire(&StrategyControl->buffer_strategy_lock);
-
- if (StrategyControl->firstFreeBuffer < 0)
- {
- SpinLockRelease(&StrategyControl->buffer_strategy_lock);
- break;
- }
-
- buf = GetBufferDescriptor(StrategyControl->firstFreeBuffer);
- Assert(buf->freeNext != FREENEXT_NOT_IN_LIST);
-
- /* Unconditionally remove buffer from freelist */
- StrategyControl->firstFreeBuffer = buf->freeNext;
- buf->freeNext = FREENEXT_NOT_IN_LIST;
-
- /*
- * Release the lock so someone else can access the freelist while
- * we check out this buffer.
- */
- SpinLockRelease(&StrategyControl->buffer_strategy_lock);
-
- /*
- * If the buffer is pinned or has a nonzero usage_count, we cannot
- * use it; discard it and retry. (This can only happen if VACUUM
- * put a valid buffer in the freelist and then someone else used
- * it before we got to it. It's probably impossible altogether as
- * of 8.3, but we'd better check anyway.)
- */
- local_buf_state = LockBufHdr(buf);
- if (BUF_STATE_GET_REFCOUNT(local_buf_state) == 0
- && BUF_STATE_GET_USAGECOUNT(local_buf_state) == 0)
- {
- if (strategy != NULL)
- AddBufferToRing(strategy, buf);
- *buf_state = local_buf_state;
- return buf;
- }
- UnlockBufHdr(buf, local_buf_state);
- }
- }
-
- /* Nothing on the freelist, so run the "clock sweep" algorithm */
+ /* Use the "clock sweep" algorithm to find a free buffer */
trycounter = NBuffers;
for (;;)
{
@@ -356,29 +286,6 @@ StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_r
}
}
-/*
- * StrategyFreeBuffer: put a buffer on the freelist
- */
-void
-StrategyFreeBuffer(BufferDesc *buf)
-{
- SpinLockAcquire(&StrategyControl->buffer_strategy_lock);
-
- /*
- * It is possible that we are told to put something in the freelist that
- * is already in it; don't screw up the list if so.
- */
- if (buf->freeNext == FREENEXT_NOT_IN_LIST)
- {
- buf->freeNext = StrategyControl->firstFreeBuffer;
- if (buf->freeNext < 0)
- StrategyControl->lastFreeBuffer = buf->buf_id;
- StrategyControl->firstFreeBuffer = buf->buf_id;
- }
-
- SpinLockRelease(&StrategyControl->buffer_strategy_lock);
-}
-
/*
* StrategySyncStart -- tell BgBufferSync where to start syncing
*
@@ -504,13 +411,6 @@ StrategyInitialize(bool init)
SpinLockInit(&StrategyControl->buffer_strategy_lock);
- /*
- * Grab the whole linked list of free buffers for our strategy. We
- * assume it was previously set up by BufferManagerShmemInit().
- */
- StrategyControl->firstFreeBuffer = 0;
- StrategyControl->lastFreeBuffer = NBuffers - 1;
-
/* Initialize the clock sweep pointer */
pg_atomic_init_u32(&StrategyControl->nextVictimBuffer, 0);
diff --git a/src/include/storage/buf_internals.h b/src/include/storage/buf_internals.h
index 52a71b138f7..00eade63971 100644
--- a/src/include/storage/buf_internals.h
+++ b/src/include/storage/buf_internals.h
@@ -217,8 +217,7 @@ BufMappingPartitionLockByIndex(uint32 index)
* single atomic variable. This layout allow us to do some operations in a
* single atomic operation, without actually acquiring and releasing spinlock;
* for instance, increase or decrease refcount. buf_id field never changes
- * after initialization, so does not need locking. freeNext is protected by
- * the buffer_strategy_lock not buffer header lock. The LWLock can take care
+ * after initialization, so does not need locking. The LWLock can take care
* of itself. The buffer header lock is *not* used to control access to the
* data in the buffer!
*
@@ -264,7 +263,6 @@ typedef struct BufferDesc
pg_atomic_uint32 state;
int wait_backend_pgprocno; /* backend of pin-count waiter */
- int freeNext; /* link in freelist chain */
PgAioWaitRef io_wref; /* set iff AIO is in progress */
LWLock content_lock; /* to lock access to buffer contents */
@@ -360,13 +358,6 @@ BufferDescriptorGetContentLock(const BufferDesc *bdesc)
return (LWLock *) (&bdesc->content_lock);
}
-/*
- * The freeNext field is either the index of the next freelist entry,
- * or one of these special values:
- */
-#define FREENEXT_END_OF_LIST (-1)
-#define FREENEXT_NOT_IN_LIST (-2)
-
/*
* Functions for acquiring/releasing a shared buffer header's spinlock. Do
* not apply these to local buffers!
@@ -453,7 +444,6 @@ extern void StrategyNotifyBgWriter(int bgwprocno);
extern Size StrategyShmemSize(void);
extern void StrategyInitialize(bool init);
-extern bool have_free_buffer(void);
/* buf_table.c */
extern Size BufTableShmemSize(int size);
--
2.49.0
From fdf3e231bb2bb07d29cf486a616ea8c4ae365abe Mon Sep 17 00:00:00 2001
From: Greg Burd <g...@burd.me>
Date: Fri, 11 Jul 2025 09:05:45 -0400
Subject: [PATCH v11 2/3] Remove the buffer_strategy_lock and make the clock
hand a 64 bit atomic
Change nextVictimBuffer to an atomic uint64 and simply atomically
increment it by 1 at each tick. The next victim buffer is the the value
of nextVictimBuffer modulo the number of buffers (NBuffers). The number
of complete passes of the clock-sweep hand is nextVictimBuffer divided
by NBuffers. Wrap-around of nextVictimBuffer would require 10 years at
~59 billion ticks per-second without restart, so unlikely that we ignore
that case entirely.
With the removal of the freelist and completePasses none of remaining
items in the BufferStrategyControl structure require strict coordination
and so it is possible to eliminate the buffer_strategy_lock as well.
---
src/backend/storage/buffer/README | 48 ++++-----
src/backend/storage/buffer/bufmgr.c | 20 +++-
src/backend/storage/buffer/freelist.c | 139 ++++++--------------------
src/backend/storage/buffer/localbuf.c | 2 +-
src/include/storage/buf_internals.h | 4 +-
5 files changed, 71 insertions(+), 142 deletions(-)
diff --git a/src/backend/storage/buffer/README b/src/backend/storage/buffer/README
index cd52effd911..d1ab222eeb8 100644
--- a/src/backend/storage/buffer/README
+++ b/src/backend/storage/buffer/README
@@ -127,11 +127,10 @@ bits of the tag's hash value. The rules stated above apply to each partition
independently. If it is necessary to lock more than one partition at a time,
they must be locked in partition-number order to avoid risk of deadlock.
-* A separate system-wide spinlock, buffer_strategy_lock, provides mutual
-exclusion for operations that select buffers for replacement. A spinlock is
-used here rather than a lightweight lock for efficiency; no other locks of any
-sort should be acquired while buffer_strategy_lock is held. This is essential
-to allow buffer replacement to happen in multiple backends with reasonable
+* Operations that select buffers for replacement don't require a lock, but
+rather use atomic operations to ensure coordination across backends when
+accessing members of the BufferStrategyControl datastructure. This allows
+buffer replacement to happen in multiple backends with reasonable
concurrency.
* Each buffer header contains a spinlock that must be taken when examining
@@ -158,9 +157,9 @@ unset by sleeping on the buffer's condition variable.
Normal Buffer Replacement Strategy
----------------------------------
-To choose a victim buffer to recycle when there are no free buffers available,
-we use a simple clock-sweep algorithm, which avoids the need to take
-system-wide locks during common operations. It works like this:
+To choose a victim buffer to recycle we use a simple clock-sweep algorithm,
+which avoids the need to take system-wide locks during common operations. It
+works like this:
Each buffer header contains a usage counter, which is incremented (up to a
small limit value) whenever the buffer is pinned. (This requires only the
@@ -168,19 +167,17 @@ buffer header spinlock, which would have to be taken anyway to increment the
buffer reference count, so it's nearly free.)
The "clock hand" is a buffer index, nextVictimBuffer, that moves circularly
-through all the available buffers. nextVictimBuffer is protected by the
-buffer_strategy_lock.
+through all the available buffers. nextVictimBuffer and completePasses are
+atomic values.
The algorithm for a process that needs to obtain a victim buffer is:
-1. Obtain buffer_strategy_lock.
+1. Select the buffer pointed to by nextVictimBuffer, and circularly advance
+nextVictimBuffer for next time.
-2. Select the buffer pointed to by nextVictimBuffer, and circularly advance
-nextVictimBuffer for next time. Release buffer_strategy_lock.
-
-3. If the selected buffer is pinned or has a nonzero usage count, it cannot
-be used. Decrement its usage count (if nonzero), reacquire
-buffer_strategy_lock, and return to step 3 to examine the next buffer.
+2. If the selected buffer is pinned or has a nonzero usage count, it cannot be
+used. Decrement its usage count (if nonzero), return to step 3 to examine the
+next buffer.
4. Pin the selected buffer, and return.
@@ -196,9 +193,9 @@ Buffer Ring Replacement Strategy
When running a query that needs to access a large number of pages just once,
such as VACUUM or a large sequential scan, a different strategy is used.
A page that has been touched only by such a scan is unlikely to be needed
-again soon, so instead of running the normal clock sweep algorithm and
+again soon, so instead of running the normal clock-sweep algorithm and
blowing out the entire buffer cache, a small ring of buffers is allocated
-using the normal clock sweep algorithm and those buffers are reused for the
+using the normal clock-sweep algorithm and those buffers are reused for the
whole scan. This also implies that much of the write traffic caused by such
a statement will be done by the backend itself and not pushed off onto other
processes.
@@ -244,13 +241,12 @@ nextVictimBuffer (which it does not change!), looking for buffers that are
dirty and not pinned nor marked with a positive usage count. It pins,
writes, and releases any such buffer.
-If we can assume that reading nextVictimBuffer is an atomic action, then
-the writer doesn't even need to take buffer_strategy_lock in order to look
-for buffers to write; it needs only to spinlock each buffer header for long
-enough to check the dirtybit. Even without that assumption, the writer
-only needs to take the lock long enough to read the variable value, not
-while scanning the buffers. (This is a very substantial improvement in
-the contention cost of the writer compared to PG 8.0.)
+We enforce reading nextVictimBuffer within an atomic action so it needs only to
+spinlock each buffer header for long enough to check the dirtybit. Even
+without that assumption, the writer only needs to take the lock long enough to
+read the variable value, not while scanning the buffers. (This is a very
+substantial improvement in the contention cost of the writer compared to PG
+8.0.)
The background writer takes shared content lock on a buffer while writing it
out (and anyone else who flushes buffer contents to disk must do so too).
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index af5ef025229..09d054a616f 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -3593,7 +3593,7 @@ BufferSync(int flags)
* This is called periodically by the background writer process.
*
* Returns true if it's appropriate for the bgwriter process to go into
- * low-power hibernation mode. (This happens if the strategy clock sweep
+ * low-power hibernation mode. (This happens if the strategy clock-sweep
* has been "lapped" and no buffer allocations have occurred recently,
* or if the bgwriter has been effectively disabled by setting
* bgwriter_lru_maxpages to 0.)
@@ -3643,7 +3643,7 @@ BgBufferSync(WritebackContext *wb_context)
uint32 new_recent_alloc;
/*
- * Find out where the clock sweep currently is, and how many buffer
+ * Find out where the clock-sweep currently is, and how many buffer
* allocations have happened since our last call.
*/
strategy_buf_id = StrategySyncStart(&strategy_passes, &recent_alloc);
@@ -3664,15 +3664,25 @@ BgBufferSync(WritebackContext *wb_context)
/*
* Compute strategy_delta = how many buffers have been scanned by the
- * clock sweep since last time. If first time through, assume none. Then
- * see if we are still ahead of the clock sweep, and if so, how many
+ * clock-sweep since last time. If first time through, assume none. Then
+ * see if we are still ahead of the clock-sweep, and if so, how many
* buffers we could scan before we'd catch up with it and "lap" it. Note:
* weird-looking coding of xxx_passes comparisons are to avoid bogus
* behavior when the passes counts wrap around.
*/
if (saved_info_valid)
{
- int32 passes_delta = strategy_passes - prev_strategy_passes;
+ int32 passes_delta;
+
+ /*
+ * It is highly unlikely that the uint64 hand of the clock-sweep
+ * would ever wrap, that would roughtly require 10 years of
+ * continuous operation at ~59 billion ticks per-second without
+ * restart.
+ */
+ Assert(prev_strategy_passes <= strategy_passes);
+
+ passes_delta = strategy_passes - prev_strategy_passes;
strategy_delta = strategy_buf_id - prev_strategy_buf_id;
strategy_delta += (long) passes_delta * NBuffers;
diff --git a/src/backend/storage/buffer/freelist.c b/src/backend/storage/buffer/freelist.c
index 162c140fb9d..906b35be4c1 100644
--- a/src/backend/storage/buffer/freelist.c
+++ b/src/backend/storage/buffer/freelist.c
@@ -27,23 +27,20 @@
/*
* The shared freelist control information.
*/
-typedef struct
-{
- /* Spinlock: protects the values below */
- slock_t buffer_strategy_lock;
-
+typedef struct {
/*
- * Clock sweep hand: index of next buffer to consider grabbing. Note that
- * this isn't a concrete buffer - we only ever increase the value. So, to
- * get an actual buffer, it needs to be used modulo NBuffers.
+ * The clock-sweep hand is atomically updated by 1 at every tick. Use the
+ * macro CLOCK_HAND_POSITION() o find the next victim's index in the
+ * BufferDescriptor array. To calculate the number of times the clock-sweep
+ * hand has made a complete pass through all available buffers in the pool
+ * divide NBuffers.
*/
- pg_atomic_uint32 nextVictimBuffer;
+ pg_atomic_uint64 nextVictimBuffer;
/*
* Statistics. These counters should be wide enough that they can't
* overflow during a single bgwriter cycle.
*/
- uint32 completePasses; /* Complete cycles of the clock sweep */
pg_atomic_uint32 numBufferAllocs; /* Buffers allocated since last reset */
/*
@@ -83,13 +80,15 @@ typedef struct BufferAccessStrategyData
Buffer buffers[FLEXIBLE_ARRAY_MEMBER];
} BufferAccessStrategyData;
-
/* Prototypes for internal functions */
static BufferDesc *GetBufferFromRing(BufferAccessStrategy strategy,
uint32 *buf_state);
static void AddBufferToRing(BufferAccessStrategy strategy,
BufferDesc *buf);
+#define CLOCK_HAND_POSITION(counter) \
+ ((counter) & 0xFFFFFFFF) % NBuffers
+
/*
* ClockSweepTick - Helper routine for StrategyGetBuffer()
*
@@ -99,6 +98,7 @@ static void AddBufferToRing(BufferAccessStrategy strategy,
static inline uint32
ClockSweepTick(void)
{
+ uint64 hand = UINT64_MAX;
uint32 victim;
/*
@@ -106,71 +106,14 @@ ClockSweepTick(void)
* doing this, this can lead to buffers being returned slightly out of
* apparent order.
*/
- victim =
- pg_atomic_fetch_add_u32(&StrategyControl->nextVictimBuffer, 1);
-
- if (victim >= NBuffers)
- {
- uint32 originalVictim = victim;
-
- /* always wrap what we look up in BufferDescriptors */
- victim = victim % NBuffers;
+ hand = pg_atomic_fetch_add_u64(&StrategyControl->nextVictimBuffer, 1);
- /*
- * If we're the one that just caused a wraparound, force
- * completePasses to be incremented while holding the spinlock. We
- * need the spinlock so StrategySyncStart() can return a consistent
- * value consisting of nextVictimBuffer and completePasses.
- */
- if (victim == 0)
- {
- uint32 expected;
- uint32 wrapped;
- bool success = false;
+ victim = CLOCK_HAND_POSITION(hand);
+ Assert(victim < NBuffers);
- expected = originalVictim + 1;
-
- while (!success)
- {
- /*
- * Acquire the spinlock while increasing completePasses. That
- * allows other readers to read nextVictimBuffer and
- * completePasses in a consistent manner which is required for
- * StrategySyncStart(). In theory delaying the increment
- * could lead to an overflow of nextVictimBuffers, but that's
- * highly unlikely and wouldn't be particularly harmful.
- */
- SpinLockAcquire(&StrategyControl->buffer_strategy_lock);
-
- wrapped = expected % NBuffers;
-
- success = pg_atomic_compare_exchange_u32(&StrategyControl->nextVictimBuffer,
- &expected, wrapped);
- if (success)
- StrategyControl->completePasses++;
- SpinLockRelease(&StrategyControl->buffer_strategy_lock);
- }
- }
- }
return victim;
}
-/*
- * have_free_buffer -- check if we've filled the buffer pool at startup
- *
- * Used exclusively by autoprewarm.
- */
-bool
-have_free_buffer(void)
-{
- uint64 hand = pg_atomic_read_u64(&StrategyControl->nextVictimBuffer);
-
- if (hand < NBuffers)
- return true;
- else
- return false;
-}
-
/*
* StrategyGetBuffer
*
@@ -193,10 +136,7 @@ StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_r
*from_ring = false;
- /*
- * If given a strategy object, see whether it can select a buffer. We
- * assume strategy objects don't need buffer_strategy_lock.
- */
+ /* If given a strategy object, see whether it can select a buffer */
if (strategy != NULL)
{
buf = GetBufferFromRing(strategy, buf_state);
@@ -241,7 +181,7 @@ StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_r
*/
pg_atomic_fetch_add_u32(&StrategyControl->numBufferAllocs, 1);
- /* Use the "clock sweep" algorithm to find a free buffer */
+ /* Use the "clock-sweep" algorithm to find a free buffer */
trycounter = NBuffers;
for (;;)
{
@@ -292,37 +232,30 @@ StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_r
* The result is the buffer index of the best buffer to sync first.
* BgBufferSync() will proceed circularly around the buffer array from there.
*
- * In addition, we return the completed-pass count (which is effectively
- * the higher-order bits of nextVictimBuffer) and the count of recent buffer
- * allocs if non-NULL pointers are passed. The alloc count is reset after
- * being read.
+ * In addition, we return the completed-pass count and the count of recent
+ * buffer allocs if non-NULL pointers are passed. The alloc count is reset
+ * after being read.
*/
-int
-StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc)
-{
- uint32 nextVictimBuffer;
- int result;
+uint32 StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc) {
+ uint64 counter = UINT64_MAX; uint32 result;
- SpinLockAcquire(&StrategyControl->buffer_strategy_lock);
- nextVictimBuffer = pg_atomic_read_u32(&StrategyControl->nextVictimBuffer);
- result = nextVictimBuffer % NBuffers;
+ counter = pg_atomic_read_u64(&StrategyControl->nextVictimBuffer);
+ result = CLOCK_HAND_POSITION(counter);
if (complete_passes)
{
- *complete_passes = StrategyControl->completePasses;
-
/*
- * Additionally add the number of wraparounds that happened before
- * completePasses could be incremented. C.f. ClockSweepTick().
+ * The number of complete passes is the counter divided by NBuffers
+ * because the clock hand is a 64-bit counter that only increases.
*/
- *complete_passes += nextVictimBuffer / NBuffers;
+ *complete_passes = (uint32) (counter / NBuffers);
}
if (num_buf_alloc)
{
*num_buf_alloc = pg_atomic_exchange_u32(&StrategyControl->numBufferAllocs, 0);
}
- SpinLockRelease(&StrategyControl->buffer_strategy_lock);
+
return result;
}
@@ -337,21 +270,14 @@ StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc)
void
StrategyNotifyBgWriter(int bgwprocno)
{
- /*
- * We acquire buffer_strategy_lock just to ensure that the store appears
- * atomic to StrategyGetBuffer. The bgwriter should call this rather
- * infrequently, so there's no performance penalty from being safe.
- */
- SpinLockAcquire(&StrategyControl->buffer_strategy_lock);
StrategyControl->bgwprocno = bgwprocno;
- SpinLockRelease(&StrategyControl->buffer_strategy_lock);
}
/*
* StrategyShmemSize
*
- * estimate the size of shared memory used by the freelist-related structures.
+ * Estimate the size of shared memory used by the freelist-related structures.
*
* Note: for somewhat historical reasons, the buffer lookup hashtable size
* is also determined here.
@@ -409,13 +335,10 @@ StrategyInitialize(bool init)
*/
Assert(init);
- SpinLockInit(&StrategyControl->buffer_strategy_lock);
-
- /* Initialize the clock sweep pointer */
- pg_atomic_init_u32(&StrategyControl->nextVictimBuffer, 0);
+ /* Initialize combined clock-sweep pointer/complete passes counter */
+ pg_atomic_init_u64(&StrategyControl->nextVictimBuffer, 0);
/* Clear statistics */
- StrategyControl->completePasses = 0;
pg_atomic_init_u32(&StrategyControl->numBufferAllocs, 0);
/* No pending notification */
@@ -659,7 +582,7 @@ GetBufferFromRing(BufferAccessStrategy strategy, uint32 *buf_state)
*
* If usage_count is 0 or 1 then the buffer is fair game (we expect 1,
* since our own previous usage of the ring element would have left it
- * there, but it might've been decremented by clock sweep since then). A
+ * there, but it might've been decremented by clock-sweep since then). A
* higher usage_count indicates someone else has touched the buffer, so we
* shouldn't re-use it.
*/
diff --git a/src/backend/storage/buffer/localbuf.c b/src/backend/storage/buffer/localbuf.c
index 3da9c41ee1d..7a34f5e430a 100644
--- a/src/backend/storage/buffer/localbuf.c
+++ b/src/backend/storage/buffer/localbuf.c
@@ -229,7 +229,7 @@ GetLocalVictimBuffer(void)
ResourceOwnerEnlarge(CurrentResourceOwner);
/*
- * Need to get a new buffer. We use a clock sweep algorithm (essentially
+ * Need to get a new buffer. We use a clock-sweep algorithm (essentially
* the same as what freelist.c does now...)
*/
trycounter = NLocBuffer;
diff --git a/src/include/storage/buf_internals.h b/src/include/storage/buf_internals.h
index 00eade63971..97002acb757 100644
--- a/src/include/storage/buf_internals.h
+++ b/src/include/storage/buf_internals.h
@@ -81,7 +81,7 @@ StaticAssertDecl(BUF_REFCOUNT_BITS + BUF_USAGECOUNT_BITS + BUF_FLAG_BITS == 32,
* accuracy and speed of the clock-sweep buffer management algorithm. A
* large value (comparable to NBuffers) would approximate LRU semantics.
* But it can take as many as BM_MAX_USAGE_COUNT+1 complete cycles of
- * clock sweeps to find a free buffer, so in practice we don't want the
+ * clock-sweeps to find a free buffer, so in practice we don't want the
* value to be very large.
*/
#define BM_MAX_USAGE_COUNT 5
@@ -439,7 +439,7 @@ extern void StrategyFreeBuffer(BufferDesc *buf);
extern bool StrategyRejectBuffer(BufferAccessStrategy strategy,
BufferDesc *buf, bool from_ring);
-extern int StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc);
+extern uint32 StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc);
extern void StrategyNotifyBgWriter(int bgwprocno);
extern Size StrategyShmemSize(void);
--
2.49.0
From 3c01069c5e51cb810949289cbe7833b10632c448 Mon Sep 17 00:00:00 2001
From: Greg Burd <g...@burd.me>
Date: Fri, 25 Jul 2025 11:53:10 -0400
Subject: [PATCH v11 3/3] Abstract clock-sweep buffer replacement algorithm
Re-author the clock-sweep algorithm such that it maintains its own state
and has a well defined API.
---
src/backend/storage/buffer/README | 20 +++---
src/backend/storage/buffer/freelist.c | 88 ++++++++++++++-------------
src/tools/pgindent/typedefs.list | 1 +
3 files changed, 57 insertions(+), 52 deletions(-)
diff --git a/src/backend/storage/buffer/README b/src/backend/storage/buffer/README
index d1ab222eeb8..3f31d04d572 100644
--- a/src/backend/storage/buffer/README
+++ b/src/backend/storage/buffer/README
@@ -166,14 +166,13 @@ small limit value) whenever the buffer is pinned. (This requires only the
buffer header spinlock, which would have to be taken anyway to increment the
buffer reference count, so it's nearly free.)
-The "clock hand" is a buffer index, nextVictimBuffer, that moves circularly
-through all the available buffers. nextVictimBuffer and completePasses are
-atomic values.
+The "clock hand" is a buffer index that moves circularly through all the
+available buffers.
The algorithm for a process that needs to obtain a victim buffer is:
-1. Select the buffer pointed to by nextVictimBuffer, and circularly advance
-nextVictimBuffer for next time.
+1. Select the buffer pointed to by the clock hand, and circularly advance it
+for next time.
2. If the selected buffer is pinned or has a nonzero usage count, it cannot be
used. Decrement its usage count (if nonzero), return to step 3 to examine the
@@ -235,13 +234,12 @@ Background Writer's Processing
------------------------------
The background writer is designed to write out pages that are likely to be
-recycled soon, thereby offloading the writing work from active backends.
-To do this, it scans forward circularly from the current position of
-nextVictimBuffer (which it does not change!), looking for buffers that are
-dirty and not pinned nor marked with a positive usage count. It pins,
-writes, and releases any such buffer.
+recycled soon, thereby offloading the writing work from active backends. To do
+this, it scans forward circularly from the current position of clock (which it
+does not change!), looking for buffers that are dirty and not pinned nor marked
+with a positive usage count. It pins, writes, and releases any such buffer.
-We enforce reading nextVictimBuffer within an atomic action so it needs only to
+We enforce reading the clock hand within an atomic action so it needs only to
spinlock each buffer header for long enough to check the dirtybit. Even
without that assumption, the writer only needs to take the lock long enough to
read the variable value, not while scanning the buffers. (This is a very
diff --git a/src/backend/storage/buffer/freelist.c b/src/backend/storage/buffer/freelist.c
index 906b35be4c1..71839dfdee9 100644
--- a/src/backend/storage/buffer/freelist.c
+++ b/src/backend/storage/buffer/freelist.c
@@ -23,19 +23,22 @@
#define INT_ACCESS_ONCE(var) ((int)(*((volatile int *)&(var))))
+typedef struct ClockSweep
+{
+ pg_atomic_uint64 counter; /* Only incremented by one */
+ uint32_t size; /* Size of the clock */
+} ClockSweep;
/*
* The shared freelist control information.
*/
-typedef struct {
+typedef struct
+{
/*
- * The clock-sweep hand is atomically updated by 1 at every tick. Use the
- * macro CLOCK_HAND_POSITION() o find the next victim's index in the
- * BufferDescriptor array. To calculate the number of times the clock-sweep
- * hand has made a complete pass through all available buffers in the pool
- * divide NBuffers.
+ * The next buffer available for use is determined by the clock-sweep
+ * algorithm.
*/
- pg_atomic_uint64 nextVictimBuffer;
+ ClockSweep clock;
/*
* Statistics. These counters should be wide enough that they can't
@@ -86,32 +89,40 @@ static BufferDesc *GetBufferFromRing(BufferAccessStrategy strategy,
static void AddBufferToRing(BufferAccessStrategy strategy,
BufferDesc *buf);
-#define CLOCK_HAND_POSITION(counter) \
- ((counter) & 0xFFFFFFFF) % NBuffers
+static void
+ClockSweepInit(ClockSweep *sweep, uint32 size)
+{
+ pg_atomic_init_u64(&sweep->counter, 0);
+ sweep->size = size;
+}
-/*
- * ClockSweepTick - Helper routine for StrategyGetBuffer()
- *
- * Move the clock hand one buffer ahead of its current position and return the
- * id of the buffer now under the hand.
- */
+/* Extract the number of complete cycles from the clock hand */
static inline uint32
-ClockSweepTick(void)
+ClockSweepCycles(ClockSweep *sweep)
{
- uint64 hand = UINT64_MAX;
- uint32 victim;
+ uint64 current = pg_atomic_read_u64(&sweep->counter);
- /*
- * Atomically move hand ahead one buffer - if there's several processes
- * doing this, this can lead to buffers being returned slightly out of
- * apparent order.
- */
- hand = pg_atomic_fetch_add_u64(&StrategyControl->nextVictimBuffer, 1);
+ return current / sweep->size;
+}
+
+/* Return the current position of the clock's hand modulo size */
+static inline uint32
+ClockSweepPosition(ClockSweep *sweep)
+{
+ uint64 counter = pg_atomic_read_u64(&sweep->counter);
+
+ return ((counter) & 0xFFFFFFFF) % sweep->size;
+}
- victim = CLOCK_HAND_POSITION(hand);
- Assert(victim < NBuffers);
+/*
+ * Move the clock hand ahead one and return its new position.
+ */
+static inline uint32
+ClockSweepTick(ClockSweep *sweep)
+{
+ uint64 counter = pg_atomic_fetch_add_u64(&sweep->counter, 1);
- return victim;
+ return ((counter) & 0xFFFFFFFF) % sweep->size;
}
/*
@@ -181,11 +192,11 @@ StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_r
*/
pg_atomic_fetch_add_u32(&StrategyControl->numBufferAllocs, 1);
- /* Use the "clock-sweep" algorithm to find a free buffer */
+ /* Use the clock-sweep algorithm to find a free buffer */
trycounter = NBuffers;
for (;;)
{
- buf = GetBufferDescriptor(ClockSweepTick());
+ buf = GetBufferDescriptor(ClockSweepTick(&StrategyControl->clock));
/*
* If the buffer is pinned or has a nonzero usage_count, we cannot use
@@ -236,19 +247,14 @@ StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_r
* buffer allocs if non-NULL pointers are passed. The alloc count is reset
* after being read.
*/
-uint32 StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc) {
- uint64 counter = UINT64_MAX; uint32 result;
-
- counter = pg_atomic_read_u64(&StrategyControl->nextVictimBuffer);
- result = CLOCK_HAND_POSITION(counter);
+uint32
+StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc)
+{
+ uint32 result = ClockSweepPosition(&StrategyControl->clock);
if (complete_passes)
{
- /*
- * The number of complete passes is the counter divided by NBuffers
- * because the clock hand is a 64-bit counter that only increases.
- */
- *complete_passes = (uint32) (counter / NBuffers);
+ *complete_passes = ClockSweepCycles(&StrategyControl->clock);
}
if (num_buf_alloc)
@@ -335,8 +341,8 @@ StrategyInitialize(bool init)
*/
Assert(init);
- /* Initialize combined clock-sweep pointer/complete passes counter */
- pg_atomic_init_u64(&StrategyControl->nextVictimBuffer, 0);
+ /* Initialize the clock-sweep algorithm */
+ ClockSweepInit(&StrategyControl->clock, NBuffers);
/* Clear statistics */
pg_atomic_init_u32(&StrategyControl->numBufferAllocs, 0);
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 3daba26b237..3ac88067e21 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -426,6 +426,7 @@ ClientCertName
ClientConnectionInfo
ClientData
ClientSocket
+ClockSweep
ClonePtrType
ClosePortalStmt
ClosePtrType
--
2.49.0
