agora inbox for [email protected]  
help / color / mirror / Atom feed
[PATCH v2 3/3] bufmgr: aio: Prototype for not waiting for already-in-progress IO
235+ messages / 2 participants
[nested] [flat]

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v2 3/3] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 150 ++++++++++++++++++++++++++--
 2 files changed, 142 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fe470de63f2..7b245daaf4e 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,46 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+/*
+ * Check if the buffer is already undergoing read AIO. If it is, assign the
+ * IO's wait reference to operation->io_wref, thereby allowing the caller to
+ * wait for that IO.
+ */
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if ((buf_state & BM_IO_IN_PROGRESS) && !(buf_state & BM_VALID))
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if ((buf_state & BM_IO_IN_PROGRESS) && !(buf_state & BM_VALID))
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1729,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1748,66 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (unlikely(operation->foreign_io))
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+
+					/*
+					 * Report and track this as a 'hit' for this backend, even
+					 * though it must have started out as a miss in
+					 * PinBufferForBlock(). The other backend (or ourselves,
+					 * as part of a read started earlier) will track this as a
+					 * 'read'.
+					 */
+					TRACE_POSTGRESQL_BUFFER_READ_DONE(operation->forknum,
+													  operation->blocknum + operation->nblocks_done,
+													  operation->smgr->smgr_rlocator.locator.spcOid,
+													  operation->smgr->smgr_rlocator.locator.dbOid,
+													  operation->smgr->smgr_rlocator.locator.relNumber,
+													  operation->smgr->smgr_rlocator.backend,
+													  true);
+
+					if (BufferIsLocal(buffer))
+						pgBufferUsage.local_blks_hit += 1;
+					else
+						pgBufferUsage.shared_blks_hit += 1;
+
+					if (operation->rel)
+						pgstat_count_buffer_hit(operation->rel);
+
+					pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+
+					if (VacuumCostActive)
+						VacuumCostBalance += VacuumCostPageHit;
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1893,43 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 * the readahead distance. This can happen e.g. for the table accesses of
+	 * an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 *
+	 * It's possible that another backend starts IO on the buffer between this
+	 * check and the ReadBuffersCanStartIO(nowait = false) below. In that case
+	 * we will synchronously wait for the IO below, but the window for that is
+	 * small enough that it won't happen often enough to have a significant
+	 * performance impact.
+	 */
+	if (ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		*nblocks_progress = 1;
+		operation->foreign_io = true;
+
+		CheckReadBuffersOperation(operation, false);
+
+
+		return true;
+	}
+
+	operation->foreign_io = false;
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1987,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O is in progress in another backend (it can't be
+	 * this backend), we want to wait for the outcome: either done, or
+	 * something went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
-- 
2.48.1.76.g4e746b1a31.dirty


--aknbaqiolq3mrqjk--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread

* [PATCH v53 7/7] Reserve replication slots specifically for REPACK
@ 2026-04-01 17:54 Álvaro Herrera <[email protected]>
  0 siblings, 0 replies; 235+ messages in thread

From: Álvaro Herrera @ 2026-04-01 17:54 UTC (permalink / raw)

This allows REPACK to not interfere with other operations that use
replication slots.  This eases configurability.
---
 doc/src/sgml/config.sgml                      | 16 ++++
 doc/src/sgml/ref/repack.sgml                  |  6 +-
 src/backend/commands/repack.c                 |  2 +-
 src/backend/commands/repack_worker.c          |  7 +-
 src/backend/replication/logical/launcher.c    |  2 +-
 src/backend/replication/logical/logical.c     |  8 +-
 .../replication/logical/logicalfuncs.c        |  2 +-
 src/backend/replication/logical/slotsync.c    |  5 +-
 src/backend/replication/slot.c                | 86 ++++++++++++-------
 src/backend/replication/slotfuncs.c           | 19 ++--
 src/backend/replication/walsender.c           |  9 +-
 src/backend/utils/misc/guc_parameters.dat     |  8 ++
 src/backend/utils/misc/postgresql.conf.sample |  2 +
 src/include/replication/logical.h             |  3 +-
 src/include/replication/slot.h                |  5 +-
 15 files changed, 117 insertions(+), 63 deletions(-)

diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index d3fea738ca3..a90d163e416 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -4638,6 +4638,22 @@ restore_command = 'copy "C:\\server\\archivedir\\%f" "%p"'  # Windows
        </listitem>
       </varlistentry>
 
+      <varlistentry id="guc-max-repack-replication-slots" xreflabel="max_repack_replication_slots">
+       <term><varname>max_repack_replication_slots</varname> (<type>integer</type>)
+       <indexterm>
+        <primary><varname>max_repack_replication_slots</varname> configuration parameter</primary>
+       </indexterm>
+       </term>
+       <listitem>
+        <para>
+         Specifies the maximum number of replication slots for use of
+         the <command>REPACK</command> command.  The default is 5.
+         This parameter can only be set at server start.
+        </para>
+       </listitem>
+      </varlistentry>
+
+
       <varlistentry id="guc-max-replication-slots" xreflabel="max_replication_slots">
        <term><varname>max_replication_slots</varname> (<type>integer</type>)
        <indexterm>
diff --git a/doc/src/sgml/ref/repack.sgml b/doc/src/sgml/ref/repack.sgml
index e993dfb3108..c532a39ee07 100644
--- a/doc/src/sgml/ref/repack.sgml
+++ b/doc/src/sgml/ref/repack.sgml
@@ -293,9 +293,9 @@ REPACK [ ( <replaceable class="parameter">option</replaceable> [, ...] ) ] USING
 
        <listitem>
         <para>
-         The <link linkend="guc-max-replication-slots"><varname>max_replication_slots</varname></link>
-         configuration parameter does not allow for creation of an additional
-         replication slot.
+         The <link linkend="guc-max-repack-replication-slots"><varname>max_repack_replication_slots</varname></link>
+         configuration parameter does not allow for the creation of an
+         additional replication slot.
         </para>
        </listitem>
       </itemizedlist>
diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c
index 92759f33969..6ba86384312 100644
--- a/src/backend/commands/repack.c
+++ b/src/backend/commands/repack.c
@@ -3381,7 +3381,7 @@ start_repack_decoding_worker(Oid relid)
 		ereport(ERROR,
 				errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
 				errmsg("out of background worker slots"),
-				errhint("You might need to increase \"%s\".", "max_worker_processes"));
+				errhint("You might need to increase \"%s\".", "max_repack_replication_slots"));
 
 	decoding_worker->seg = seg;
 	decoding_worker->error_mqh = mqh;
diff --git a/src/backend/commands/repack_worker.c b/src/backend/commands/repack_worker.c
index c85166ba849..39cbb0252f4 100644
--- a/src/backend/commands/repack_worker.c
+++ b/src/backend/commands/repack_worker.c
@@ -223,7 +223,7 @@ repack_setup_logical_decoding(Oid relid)
 	 * Make sure we can use logical decoding.
 	 */
 	CheckSlotPermissions();
-	CheckLogicalDecodingRequirements();
+	CheckLogicalDecodingRequirements(true);
 
 	/*
 	 * A single backend should not execute multiple REPACK commands at a time,
@@ -232,8 +232,8 @@ repack_setup_logical_decoding(Oid relid)
 	 * RS_TEMPORARY so that the slot gets cleaned up on ERROR.
 	 */
 	snprintf(NameStr(slotname), NAMEDATALEN, "repack_%d", MyProcPid);
-	ReplicationSlotCreate(NameStr(slotname), true, RS_TEMPORARY, false, false,
-						  false);
+	ReplicationSlotCreate(NameStr(slotname), true, RS_TEMPORARY, false, true,
+						  false, false);
 
 	EnsureLogicalDecodingEnabled();
 
@@ -244,6 +244,7 @@ repack_setup_logical_decoding(Oid relid)
 	ctx = CreateInitDecodingContext(REPL_PLUGIN_NAME,
 									NIL,
 									true,
+									true,
 									InvalidXLogRecPtr,
 									XL_ROUTINE(.page_read = read_local_xlog_page,
 											   .segment_open = wal_segment_open,
diff --git a/src/backend/replication/logical/launcher.c b/src/backend/replication/logical/launcher.c
index 09964198550..d83125afd0d 100644
--- a/src/backend/replication/logical/launcher.c
+++ b/src/backend/replication/logical/launcher.c
@@ -1575,7 +1575,7 @@ CreateConflictDetectionSlot(void)
 			errmsg("creating replication conflict detection slot"));
 
 	ReplicationSlotCreate(CONFLICT_DETECTION_SLOT, false, RS_PERSISTENT, false,
-						  false, false);
+						  false, false, false);
 
 	init_conflict_slot_xmin();
 }
diff --git a/src/backend/replication/logical/logical.c b/src/backend/replication/logical/logical.c
index dc46a372ead..06efcaa60b2 100644
--- a/src/backend/replication/logical/logical.c
+++ b/src/backend/replication/logical/logical.c
@@ -108,9 +108,9 @@ static void LoadOutputPlugin(OutputPluginCallbacks *callbacks, const char *plugi
  * decoding.
  */
 void
-CheckLogicalDecodingRequirements(void)
+CheckLogicalDecodingRequirements(bool repack)
 {
-	CheckSlotRequirements();
+	CheckSlotRequirements(repack);
 
 	/*
 	 * NB: Adding a new requirement likely means that RestoreSlotFromDisk()
@@ -308,6 +308,7 @@ StartupDecodingContext(List *output_plugin_options,
  * output_plugin_options -- contains options passed to the output plugin
  * need_full_snapshot -- if true, must obtain a snapshot able to read all
  *		tables; if false, one that can read only catalogs is acceptable.
+ * for_repack -- if true, we're going to be decoding for REPACK.
  * restart_lsn -- if given as invalid, it's this routine's responsibility to
  *		mark WAL as reserved by setting a convenient restart_lsn for the slot.
  *		Otherwise, we set for decoding to start from the given LSN without
@@ -328,6 +329,7 @@ LogicalDecodingContext *
 CreateInitDecodingContext(const char *plugin,
 						  List *output_plugin_options,
 						  bool need_full_snapshot,
+						  bool for_repack,
 						  XLogRecPtr restart_lsn,
 						  XLogReaderRoutine *xl_routine,
 						  LogicalOutputPluginWriterPrepareWrite prepare_write,
@@ -344,7 +346,7 @@ CreateInitDecodingContext(const char *plugin,
 	 * On a standby, this check is also required while creating the slot.
 	 * Check the comments in the function.
 	 */
-	CheckLogicalDecodingRequirements();
+	CheckLogicalDecodingRequirements(for_repack);
 
 	/* shorter lines... */
 	slot = MyReplicationSlot;
diff --git a/src/backend/replication/logical/logicalfuncs.c b/src/backend/replication/logical/logicalfuncs.c
index 9760818941d..512013b0ef0 100644
--- a/src/backend/replication/logical/logicalfuncs.c
+++ b/src/backend/replication/logical/logicalfuncs.c
@@ -115,7 +115,7 @@ pg_logical_slot_get_changes_guts(FunctionCallInfo fcinfo, bool confirm, bool bin
 
 	CheckSlotPermissions();
 
-	CheckLogicalDecodingRequirements();
+	CheckLogicalDecodingRequirements(false);
 
 	if (PG_ARGISNULL(0))
 		ereport(ERROR,
diff --git a/src/backend/replication/logical/slotsync.c b/src/backend/replication/logical/slotsync.c
index e75db69e3f6..28c89c5e10d 100644
--- a/src/backend/replication/logical/slotsync.c
+++ b/src/backend/replication/logical/slotsync.c
@@ -425,7 +425,7 @@ get_local_synced_slots(void)
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
 
-	for (int i = 0; i < max_replication_slots; i++)
+	for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 
@@ -814,6 +814,7 @@ synchronize_one_slot(RemoteSlot *remote_slot, Oid remote_dbid,
 		 */
 		ReplicationSlotCreate(remote_slot->name, true, RS_TEMPORARY,
 							  remote_slot->two_phase,
+							  false,
 							  remote_slot->failover,
 							  true);
 
@@ -1691,7 +1692,7 @@ update_synced_slots_inactive_since(void)
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
 
-	for (int i = 0; i < max_replication_slots; i++)
+	for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 
diff --git a/src/backend/replication/slot.c b/src/backend/replication/slot.c
index 9533515a63b..47679161b67 100644
--- a/src/backend/replication/slot.c
+++ b/src/backend/replication/slot.c
@@ -151,6 +151,8 @@ ReplicationSlot *MyReplicationSlot = NULL;
 /* GUC variables */
 int			max_replication_slots = 10; /* the maximum number of replication
 										 * slots */
+int			max_repack_replication_slots = 5;	/* the maximum number of slots
+												 * for REPACK */
 
 /*
  * Invalidate replication slots that have remained idle longer than this
@@ -188,14 +190,15 @@ static void SaveSlotToPath(ReplicationSlot *slot, const char *dir, int elevel);
 Size
 ReplicationSlotsShmemSize(void)
 {
+	int			totalslots = max_replication_slots + max_repack_replication_slots;
 	Size		size = 0;
 
-	if (max_replication_slots == 0)
+	if (totalslots == 0)
 		return size;
 
 	size = offsetof(ReplicationSlotCtlData, replication_slots);
 	size = add_size(size,
-					mul_size(max_replication_slots, sizeof(ReplicationSlot)));
+					mul_size(totalslots, sizeof(ReplicationSlot)));
 
 	return size;
 }
@@ -208,7 +211,7 @@ ReplicationSlotsShmemInit(void)
 {
 	bool		found;
 
-	if (max_replication_slots == 0)
+	if (max_replication_slots + max_repack_replication_slots == 0)
 		return;
 
 	ReplicationSlotCtl = (ReplicationSlotCtlData *)
@@ -222,7 +225,7 @@ ReplicationSlotsShmemInit(void)
 		/* First time through, so initialize */
 		MemSet(ReplicationSlotCtl, 0, ReplicationSlotsShmemSize());
 
-		for (i = 0; i < max_replication_slots; i++)
+		for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 		{
 			ReplicationSlot *slot = &ReplicationSlotCtl->replication_slots[i];
 
@@ -372,6 +375,7 @@ IsSlotForConflictCheck(const char *name)
  * db_specific: logical decoding is db specific; if the slot is going to
  *	   be used for that pass true, otherwise false.
  * two_phase: If enabled, allows decoding of prepared transactions.
+ * repack: If true, use a slot from the pool for REPACK.
  * failover: If enabled, allows the slot to be synced to standbys so
  *     that logical replication can be resumed after failover.
  * synced: True if the slot is synchronized from the primary server.
@@ -379,10 +383,11 @@ IsSlotForConflictCheck(const char *name)
 void
 ReplicationSlotCreate(const char *name, bool db_specific,
 					  ReplicationSlotPersistency persistency,
-					  bool two_phase, bool failover, bool synced)
+					  bool two_phase, bool repack, bool failover, bool synced)
 {
 	ReplicationSlot *slot = NULL;
-	int			i;
+	int			startpoint,
+				endpoint;
 
 	Assert(MyReplicationSlot == NULL);
 
@@ -431,12 +436,16 @@ ReplicationSlotCreate(const char *name, bool db_specific,
 	LWLockAcquire(ReplicationSlotAllocationLock, LW_EXCLUSIVE);
 
 	/*
-	 * Check for name collision, and identify an allocatable slot.  We need to
-	 * hold ReplicationSlotControlLock in shared mode for this, so that nobody
-	 * else can change the in_use flags while we're looking at them.
+	 * Check for name collision (across the whole array), and identify an
+	 * allocatable slot (in the array slice specific to our current use case:
+	 * either general, or REPACK only).  We need to hold
+	 * ReplicationSlotControlLock in shared mode for this, so that nobody else
+	 * can change the in_use flags while we're looking at them.
 	 */
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (i = 0; i < max_replication_slots; i++)
+	startpoint = !repack ? 0 : max_replication_slots;
+	endpoint = max_replication_slots + (repack ? max_repack_replication_slots : 0);
+	for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 
@@ -444,7 +453,9 @@ ReplicationSlotCreate(const char *name, bool db_specific,
 			ereport(ERROR,
 					(errcode(ERRCODE_DUPLICATE_OBJECT),
 					 errmsg("replication slot \"%s\" already exists", name)));
-		if (!s->in_use && slot == NULL)
+
+		if (i >= startpoint && i < endpoint &&
+			!s->in_use && slot == NULL)
 			slot = s;
 	}
 	LWLockRelease(ReplicationSlotControlLock);
@@ -454,7 +465,8 @@ ReplicationSlotCreate(const char *name, bool db_specific,
 		ereport(ERROR,
 				(errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED),
 				 errmsg("all replication slots are in use"),
-				 errhint("Free one or increase \"max_replication_slots\".")));
+				 errhint("Free one or increase \"%s\".",
+						 repack ? "max_repack_replication_slots" : "max_replication_slots")));
 
 	/*
 	 * Since this slot is not in use, nobody should be looking at any part of
@@ -547,7 +559,7 @@ SearchNamedReplicationSlot(const char *name, bool need_lock)
 	if (need_lock)
 		LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
 
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 
@@ -575,7 +587,8 @@ int
 ReplicationSlotIndex(ReplicationSlot *slot)
 {
 	Assert(slot >= ReplicationSlotCtl->replication_slots &&
-		   slot < ReplicationSlotCtl->replication_slots + max_replication_slots);
+		   slot < ReplicationSlotCtl->replication_slots +
+		   (max_replication_slots + max_repack_replication_slots));
 
 	return slot - ReplicationSlotCtl->replication_slots;
 }
@@ -869,7 +882,7 @@ ReplicationSlotCleanup(bool synced_only)
 restart:
 	found_valid_logicalslot = false;
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 
@@ -1251,7 +1264,7 @@ ReplicationSlotsComputeRequiredXmin(bool already_locked)
 	if (!already_locked)
 		LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
 
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 		TransactionId effective_xmin;
@@ -1306,7 +1319,7 @@ ReplicationSlotsComputeRequiredLSN(void)
 	Assert(ReplicationSlotCtl != NULL);
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 		XLogRecPtr	restart_lsn;
@@ -1373,12 +1386,12 @@ ReplicationSlotsComputeLogicalRestartLSN(void)
 	XLogRecPtr	result = InvalidXLogRecPtr;
 	int			i;
 
-	if (max_replication_slots <= 0)
+	if (max_replication_slots + max_repack_replication_slots <= 0)
 		return InvalidXLogRecPtr;
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
 
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s;
 		XLogRecPtr	restart_lsn;
@@ -1453,11 +1466,11 @@ ReplicationSlotsCountDBSlots(Oid dboid, int *nslots, int *nactive)
 
 	*nslots = *nactive = 0;
 
-	if (max_replication_slots <= 0)
+	if (max_replication_slots + max_repack_replication_slots <= 0)
 		return false;
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s;
 
@@ -1514,13 +1527,13 @@ ReplicationSlotsDropDBSlots(Oid dboid)
 	bool		found_valid_logicalslot;
 	bool		dropped = false;
 
-	if (max_replication_slots <= 0)
+	if (max_replication_slots + max_repack_replication_slots <= 0)
 		return;
 
 restart:
 	found_valid_logicalslot = false;
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s;
 		char	   *slotname;
@@ -1617,11 +1630,11 @@ CheckLogicalSlotExists(void)
 {
 	bool		found = false;
 
-	if (max_replication_slots <= 0)
+	if (max_replication_slots + max_repack_replication_slots <= 0)
 		return false;
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (int i = 0; i < max_replication_slots; i++)
+	for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s;
 		bool		invalidated;
@@ -1655,17 +1668,24 @@ CheckLogicalSlotExists(void)
  * slots.
  */
 void
-CheckSlotRequirements(void)
+CheckSlotRequirements(bool repack)
 {
 	/*
 	 * NB: Adding a new requirement likely means that RestoreSlotFromDisk()
 	 * needs the same check.
 	 */
 
-	if (max_replication_slots == 0)
+	if (!repack && max_replication_slots == 0)
 		ereport(ERROR,
-				(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
-				 errmsg("replication slots can only be used if \"max_replication_slots\" > 0")));
+				errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+				errmsg("replication slots can only be used if \"%s\" > 0",
+					   "max_replication_slots"));
+
+	if (repack && max_repack_replication_slots == 0)
+		ereport(ERROR,
+				errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+				errmsg("REPACK can only be used if \"%s\" > 0",
+					   "max_repack_replication_slots"));
 
 	if (wal_level < WAL_LEVEL_REPLICA)
 		ereport(ERROR,
@@ -2216,7 +2236,7 @@ InvalidateObsoleteReplicationSlots(uint32 possible_causes,
 	Assert(!(possible_causes & RS_INVAL_WAL_REMOVED) || oldestSegno > 0);
 	Assert(possible_causes != RS_INVAL_NONE);
 
-	if (max_replication_slots == 0)
+	if (max_replication_slots == 0 && max_repack_replication_slots == 0)
 		return invalidated;
 
 	XLogSegNoOffsetToRecPtr(oldestSegno, 0, wal_segment_size, oldestLSN);
@@ -2224,7 +2244,7 @@ InvalidateObsoleteReplicationSlots(uint32 possible_causes,
 restart:
 	found_valid_logicalslot = false;
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (int i = 0; i < max_replication_slots; i++)
+	for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 		bool		released_lock = false;
@@ -2329,7 +2349,7 @@ CheckPointReplicationSlots(bool is_shutdown)
 	 */
 	LWLockAcquire(ReplicationSlotAllocationLock, LW_SHARED);
 
-	for (i = 0; i < max_replication_slots; i++)
+	for (i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 		char		path[MAXPGPATH];
@@ -2430,7 +2450,7 @@ StartupReplicationSlots(void)
 	FreeDir(replication_dir);
 
 	/* currently no slots exist, we're done. */
-	if (max_replication_slots <= 0)
+	if (max_replication_slots + max_repack_replication_slots <= 0)
 		return;
 
 	/* Now that we have recovered all the data, compute replication xmin */
diff --git a/src/backend/replication/slotfuncs.c b/src/backend/replication/slotfuncs.c
index 9f5e4f998fe..16fbd383735 100644
--- a/src/backend/replication/slotfuncs.c
+++ b/src/backend/replication/slotfuncs.c
@@ -53,7 +53,7 @@ create_physical_replication_slot(char *name, bool immediately_reserve,
 	/* acquire replication slot, this will check for conflicting names */
 	ReplicationSlotCreate(name, false,
 						  temporary ? RS_TEMPORARY : RS_PERSISTENT, false,
-						  false, false);
+						  false, false, false);
 
 	if (immediately_reserve)
 	{
@@ -90,7 +90,7 @@ pg_create_physical_replication_slot(PG_FUNCTION_ARGS)
 
 	CheckSlotPermissions();
 
-	CheckSlotRequirements();
+	CheckSlotRequirements(false);
 
 	create_physical_replication_slot(NameStr(*name),
 									 immediately_reserve,
@@ -146,7 +146,7 @@ create_logical_replication_slot(char *name, char *plugin,
 	 */
 	ReplicationSlotCreate(name, true,
 						  temporary ? RS_TEMPORARY : RS_EPHEMERAL, two_phase,
-						  failover, false);
+						  false, failover, false);
 
 	/*
 	 * Ensure the logical decoding is enabled before initializing the logical
@@ -164,6 +164,7 @@ create_logical_replication_slot(char *name, char *plugin,
 	 */
 	ctx = CreateInitDecodingContext(plugin, NIL,
 									false,	/* just catalogs is OK */
+									false,	/* not repack */
 									restart_lsn,
 									XL_ROUTINE(.page_read = read_local_xlog_page,
 											   .segment_open = wal_segment_open,
@@ -203,7 +204,7 @@ pg_create_logical_replication_slot(PG_FUNCTION_ARGS)
 
 	CheckSlotPermissions();
 
-	CheckLogicalDecodingRequirements();
+	CheckLogicalDecodingRequirements(false);
 
 	create_logical_replication_slot(NameStr(*name),
 									NameStr(*plugin),
@@ -240,7 +241,7 @@ pg_drop_replication_slot(PG_FUNCTION_ARGS)
 
 	CheckSlotPermissions();
 
-	CheckSlotRequirements();
+	CheckSlotRequirements(false);
 
 	ReplicationSlotDrop(NameStr(*name), true);
 
@@ -270,7 +271,7 @@ pg_get_replication_slots(PG_FUNCTION_ARGS)
 	currlsn = GetXLogWriteRecPtr();
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
-	for (slotno = 0; slotno < max_replication_slots; slotno++)
+	for (slotno = 0; slotno < max_replication_slots + max_repack_replication_slots; slotno++)
 	{
 		ReplicationSlot *slot = &ReplicationSlotCtl->replication_slots[slotno];
 		ReplicationSlot slot_contents;
@@ -648,9 +649,9 @@ copy_replication_slot(FunctionCallInfo fcinfo, bool logical_slot)
 	CheckSlotPermissions();
 
 	if (logical_slot)
-		CheckLogicalDecodingRequirements();
+		CheckLogicalDecodingRequirements(false);
 	else
-		CheckSlotRequirements();
+		CheckSlotRequirements(false);
 
 	LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
 
@@ -665,7 +666,7 @@ copy_replication_slot(FunctionCallInfo fcinfo, bool logical_slot)
 	 * managed to create the new slot, we advance the new slot's restart_lsn
 	 * to the source slot's updated restart_lsn the second time we lock it.
 	 */
-	for (int i = 0; i < max_replication_slots; i++)
+	for (int i = 0; i < max_replication_slots + max_repack_replication_slots; i++)
 	{
 		ReplicationSlot *s = &ReplicationSlotCtl->replication_slots[i];
 
diff --git a/src/backend/replication/walsender.c b/src/backend/replication/walsender.c
index 2bb3f34dc6d..9d7d675fa96 100644
--- a/src/backend/replication/walsender.c
+++ b/src/backend/replication/walsender.c
@@ -1220,7 +1220,7 @@ CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
 	{
 		ReplicationSlotCreate(cmd->slotname, false,
 							  cmd->temporary ? RS_TEMPORARY : RS_PERSISTENT,
-							  false, false, false);
+							  false, false, false, false);
 
 		if (reserve_wal)
 		{
@@ -1240,7 +1240,7 @@ CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
 
 		Assert(cmd->kind == REPLICATION_KIND_LOGICAL);
 
-		CheckLogicalDecodingRequirements();
+		CheckLogicalDecodingRequirements(false);
 
 		/*
 		 * Initially create persistent slot as ephemeral - that allows us to
@@ -1251,7 +1251,7 @@ CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
 		 */
 		ReplicationSlotCreate(cmd->slotname, true,
 							  cmd->temporary ? RS_TEMPORARY : RS_EPHEMERAL,
-							  two_phase, failover, false);
+							  two_phase, false, failover, false);
 
 		/*
 		 * Do options check early so that we can bail before calling the
@@ -1309,6 +1309,7 @@ CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
 		Assert(IsLogicalDecodingEnabled());
 
 		ctx = CreateInitDecodingContext(cmd->plugin, NIL, need_full_snapshot,
+										false,
 										InvalidXLogRecPtr,
 										XL_ROUTINE(.page_read = logical_read_xlog_page,
 												   .segment_open = WalSndSegmentOpen,
@@ -1466,7 +1467,7 @@ StartLogicalReplication(StartReplicationCmd *cmd)
 	QueryCompletion qc;
 
 	/* make sure that our requirements are still fulfilled */
-	CheckLogicalDecodingRequirements();
+	CheckLogicalDecodingRequirements(false);
 
 	Assert(!MyReplicationSlot);
 
diff --git a/src/backend/utils/misc/guc_parameters.dat b/src/backend/utils/misc/guc_parameters.dat
index a315c4ab8ab..bf82b8f6048 100644
--- a/src/backend/utils/misc/guc_parameters.dat
+++ b/src/backend/utils/misc/guc_parameters.dat
@@ -2071,6 +2071,14 @@
   max => 'MAX_BACKENDS',
 },
 
+{ name => 'max_repack_replication_slots', type => 'int', context => 'PGC_POSTMASTER', group => 'REPLICATION_SENDING',
+  short_desc => 'Sets the maximum number of replication slots for use by REPACK.',
+  variable => 'max_repack_replication_slots',
+  boot_val => '5',
+  min => '0',
+  max => 'MAX_BACKENDS',
+},
+
 /* see max_wal_senders */
 { name => 'max_replication_slots', type => 'int', context => 'PGC_POSTMASTER', group => 'REPLICATION_SENDING',
   short_desc => 'Sets the maximum number of simultaneously defined replication slots.',
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index 6d0337853e0..fb75990609e 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -348,6 +348,8 @@
                                 # (change requires restart)
 #max_replication_slots = 10     # max number of replication slots
                                 # (change requires restart)
+#max_repack_replication_slots = 5 # max number of replication slots for REPACK
+                                # (change requires restart)
 #wal_keep_size = 0              # in megabytes; 0 disables
 #max_slot_wal_keep_size = -1    # in megabytes; -1 disables
 #idle_replication_slot_timeout = 0      # in seconds; 0 disables
diff --git a/src/include/replication/logical.h b/src/include/replication/logical.h
index bc9d4ece672..bc075b16741 100644
--- a/src/include/replication/logical.h
+++ b/src/include/replication/logical.h
@@ -115,11 +115,12 @@ typedef struct LogicalDecodingContext
 } LogicalDecodingContext;
 
 
-extern void CheckLogicalDecodingRequirements(void);
+extern void CheckLogicalDecodingRequirements(bool repack);
 
 extern LogicalDecodingContext *CreateInitDecodingContext(const char *plugin,
 														 List *output_plugin_options,
 														 bool need_full_snapshot,
+														 bool for_repack,
 														 XLogRecPtr restart_lsn,
 														 XLogReaderRoutine *xl_routine,
 														 LogicalOutputPluginWriterPrepareWrite prepare_write,
diff --git a/src/include/replication/slot.h b/src/include/replication/slot.h
index 4b4709f6e2c..489af7d8d6c 100644
--- a/src/include/replication/slot.h
+++ b/src/include/replication/slot.h
@@ -324,6 +324,7 @@ extern PGDLLIMPORT ReplicationSlot *MyReplicationSlot;
 
 /* GUCs */
 extern PGDLLIMPORT int max_replication_slots;
+extern PGDLLIMPORT int max_repack_replication_slots;
 extern PGDLLIMPORT char *synchronized_standby_slots;
 extern PGDLLIMPORT int idle_replication_slot_timeout_secs;
 
@@ -334,7 +335,7 @@ extern void ReplicationSlotsShmemInit(void);
 /* management of individual slots */
 extern void ReplicationSlotCreate(const char *name, bool db_specific,
 								  ReplicationSlotPersistency persistency,
-								  bool two_phase, bool failover,
+								  bool two_phase, bool repack, bool failover,
 								  bool synced);
 extern void ReplicationSlotPersist(void);
 extern void ReplicationSlotDrop(const char *name, bool nowait);
@@ -377,7 +378,7 @@ extern void ReplicationSlotDropAtPubNode(WalReceiverConn *wrconn, char *slotname
 extern void StartupReplicationSlots(void);
 extern void CheckPointReplicationSlots(bool is_shutdown);
 
-extern void CheckSlotRequirements(void);
+extern void CheckSlotRequirements(bool repack);
 extern void CheckSlotPermissions(void);
 extern ReplicationSlotInvalidationCause
 			GetSlotInvalidationCause(const char *cause_name);
-- 
2.47.3


--qr3jlalmmcpkiodg--





^ permalink  raw  reply  [nested|flat] 235+ messages in thread


end of thread, other threads:[~2026-04-01 17:54 UTC | newest]

Thread overview: 235+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v2 3/3] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2025-08-15 15:01 [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO Andres Freund <[email protected]>
2026-04-01 17:54 [PATCH v53 7/7] Reserve replication slots specifically for REPACK Álvaro Herrera <[email protected]>

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox